Xml data clustering an overview

XML Data Clustering: An Overview
ALSAYED ALGERGAWY, Magdeburg University
MARCO MESITI, University of Milano
RICHI NAYAK, Queensland University of Technology
GUNTER SAAKE, Magdeburg University

In the last few years we have observed a proliferation of approaches for clustering XML documents and
schemas based on their structure and content. The presence of such a huge amount of approaches is due to
the different applications requiring the clustering of XML data. These applications need data in the form
of similar contents, tags, paths, structures, and semantics. In this article, we first outline the application
contexts in which clustering is useful, then we survey approaches so far proposed relying on the abstract
representation of data (instances or schema), on the identified similarity measure, and on the clustering algorithm. In this presentation, we aim to draw a taxonomy in which the current approaches can be classified
and compared. We aim at introducing an integrated view that is useful when comparing XML data clustering approaches, when developing a new clustering algorithm, and when implementing an XML clustering
component. Finally, the article moves into the description of future trends and research issues that still need
to be faced.
Categories and Subject Descriptors: H.2.8 [Database Management]: Database Applications—Data mining;
H.3.3 [Information Storage and Retrieval]: Information Search and Retrieval; H.3.5 [Information
Storage and Retrieval]: Online Information Services
General Terms: Documentation, Algorithms, Performance
Additional Key Words and Phrases: XML data, clustering, tree similarity, schema matching, semantic similarity, structural similarity, documentation
ACM Reference Format:
Algergawy, A., Mesiti, M., Nayak, R., and Saake, G. 2011. XML data clustering: An overview. ACM Comput.
Surv. 43, 4, Article 25 (October 2011), 41 pages.
DOI = 10.1145/1978802.1978804 http://doi.acm.org/10.1145/1978802.1978804

1. INTRODUCTION

The eXtensible Markup Language (XML) has emerged as a standard for information representation and exchange on the Web and the Internet [Wilde and Glushko
2008]. Consequently, a huge amount of information is represented in XML and several
tools have been developed to deliver, store, integrate, and query XML data [Wang et al.
2004; Bertino and Ferrari 2001; Florescu and Kossmann 1999]. It becomes inevitable to
A. Algergawy has been supported by the Egyptian Ministry of Higher Education and Tanta University,
Egypt.
A. Algergawy is currently affiliated with Tanta University, Tanta, Egypt.
Authors’ addresses: A. Algergawy, Computer Engineering Department, Tanta University, Tanta, Egypt;
email: algergawy@informatik.uni-leipzig.de; G. Saake, Magdeburg University, Computer Science Department, 39106 Magdeburg, Germany; email: saake@cs.unimagdeburg.de; M. Mesiti, DICO-Universita’ di Milano, via Comelico 39/41, 20135, Milano, Italy; email: mesiti@dico.unimi.it; R. Nayak, Queensland University of Technology, Faculty of Science and Technology, GPO Box 2434 Brisbane, Australia; email:
r.nayak@qut.edu.au.
Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted
without fee provided that copies are not made or distributed for profit or commercial advantage and that
copies show this notice on the first page or initial screen of a display along with the full citation. Copyrights for
components of this work owned by others than ACM must be honored. Abstracting with credit is permitted.
To copy otherwise, to republish, to post on servers, to redistribute to lists, or to use any component of this
work in other works requires prior specific permission and/or a fee. Permissions may be requested from
Publications Dept., ACM, Inc., 2 Penn Plaza, Suite 701, New York, NY 10121-0701 USA, fax +1 (212)
869-0481, or permissions@acm.org.
c 2011 ACM 0360-0300/2011/10-ART25 $10.00
DOI 10.1145/1978802.1978804 http://doi.acm.org/10.1145/1978802.1978804
ACM Computing Surveys, Vol. 43, No. 4, Article 25, Publication date: October 2011.

25

25:2

A. Algergawy et al.

develop high-performance techniques for efficiently managing and analyzing extremely
large collections of XML data. One of the methods that many researchers have focused
on is clustering, which groups similar XML data according to their content, and structures. The clustering process of XML data plays a crucial role in many data application
domains, such as information retrieval, data integration, document classification, Web
mining, and query processing.
Clustering in general, is a useful technique for grouping data objects within a single
group/cluster that share similar features, while placing objects in different groups
that are dissimilar [Jain et al. 1999; Berkhin 2002; Xu and Wunsch 2005]. Specific
research on clustering XML data is gaining momentum [Lee et al. 2002; Lian et al.
2004; Leung et al. 2005; Dalamagas et al. 2006; Nayak and Tran 2007; Aggarwal et al.
2007; Choi et al. 2007; Nayak 2008] both for clustering XML documents and XML
schemas according to their contents and structures. Several XML schema languages
have been proposed [Lee and Chu 2000] for the description of the structure and the
legal building blocks of an XML document. Among them, XML Data Type Definition
(DTD) and XML Schema Definition (XSD) are commonly used. Since the document
definition outlined in a schema holds true for all document instances of that schema,
the result produced from the clustering of schemas is able to group together documents
that present similar characteristics. However, in practice, some XML documents do not
have an associated schema and schema instances might present different structures
due to the employment of the choice operator. Therefore, algorithms for clustering both
XML documents and XML schemas have attracted attention from researchers.
Clustering XML data is an intricate process and it differs significantly from clustering of flat data and text. The difficulties of clustering XML data are due to the following
reasons [Aggarwal et al. 2007]. First, clustering algorithms require the computation
of similarity between different sets of XML data, which is itself a difficult research
problem (the heterogeneity in XML data presents many challenges to identify the ideal
similarity function). Second, the structural organization of XML data increases implicit
dimensionality that a clustering algorithm needs to handle, which leads to meaningless
clusters. XML data have several features, such as semantic, structure, and content,
each containing a set of subfeatures. Clustering XML data, considering one feature
while ignoring the others, fails to achieve accurate cluster results. For example, Figure 1(a) shows three XML schemas representing journal and conference papers in the
DBLP database. The data set has common elements such as “Author” and “Title.” Even
if D1 and D2 have only one different element, they should be in two different clusters
according to usual semantics that give different relevance to publications in journals
and conferences. In contrast, even if D2 and D3 have only one different element, they
should be in the same cluster because both refer to conference papers. Furthermore,
the need to organize XML documents according to their content and structure has
become challenging, due to the increase of XML data heterogeneity. Figure 1(b) depicts
the fragments of six XML documents from the publishing domain: the XML fragments
shown in (a), (b), (c) and (f) share a similar structure, and the fragments in (d) and (e)
share a similar structure. It can be observed that the fragments in (a) and (f) share a
similar structure to fragments in (b) and (c), however, these two sets of fragments differ
in their content. These documents will be grouped in two clusters about “Books” and
“Conference Articles” if structural similarity is considered as a criterion for clustering.
However, this kind of grouping will fail to further distinguish the documents in the
“Books” cluster that contains books of several genres. On the other hand, clustering of
documents based only on content features similarity will fail to distinguish between
conference articles and books that follow two different structures. In order to derive a
meaningful grouping, these fragments should be analyzed in terms of both their structural and content feature similarity. Clustering the XML documents by considering the


25:3

(a) A set of XML schemas.

(b) A set of XML documents.
Fig. 1. Examples of XML data.

structural and content features together will result in three clusters, namely “Books
on Data Mining (DM),” “Books on Biology (Bio),” and “Conference articles on Data
Mining.”
In order to conduct a good survey and to construct a fair base for comparing existing
XML data clustering approaches, a high-level architecture for a generic framework of
XML data clustering is proposed. Inspired by data clustering activity steps [Jain et al.
1999], Figure 2 depicts the framework of XML data clustering with three basic phases.

25:4

A. Algergawy et al.

Fig. 2. A Generic XML data clustering framework.

(1) Data representation. XML data are represented using a common data model that
captures semantic and structure information inherent in XML data. This phase
includes two subphases: feature selection and feature extraction. Feature selection chooses distinctive features from a set of candidates, while feature extraction
employs rules to generate useful and novel features from the original ones. We
elaborate on XML data representation in Section 3.
(2) Similarity computation. The proximity functions to measure the similarity between
pairs of data objects are determined. XML data are grouped according to the similarity of the extracted/selected features. Performance of the clustering solution
mainly depends upon the similarity measure employed. Based on the type of data
model used to represent XML data, several XML similarity measures have been
proposed. We discuss XML similarity measures in Section 4.
(3) Clustering/grouping. Similar XML data are grouped together based on a proximity
function using a proper clustering algorithm. The majority of clustering algorithms
are implicitly or explicitly linked to the similarity measures employed. In Figure 2,
the thin arrows between the “similarity computation” and the “clustering/grouping”
boxes indicate that the grouping process can be interleaved with the similarity computation phase. Finally, the output of the clustering framework can be represented
either as a set of clusters or as nested sets of data (hierarchies), depicted as dotted
lines in Figure 2. We make a detailed discussion on the clustering approaches and
the evaluation of the quality of their application in Section 5.
This article presents an overview of XML data clustering methodologies and implementations in order to draw a road map of using clustering algorithms in XML data
management. The article starts from the application contexts where XML data clustering is useful, and then surveys the current approaches, and presents a taxonomy that
explains their common features. The article also includes a discussion of the challenges
and beneﬁts that the ﬁeld of XML data clustering brings forward. It is hoped that the
survey will be helpful both to developers of new approaches and to users who need to
select a method from a library of available approaches.


25:5

The remainder of the article is organized as follows. Section 2 surveys different
application domains that utilize the output of XML data clustering. The three phases
of the generic clustering framework are discussed in Sections 3, 4, and 5, respectively.
In Section 6, the existing approaches are presented and compared according to the
introduced framework. Concluding remarks and open research directions are presented
in Section 7.
2. APPLICATION DOMAINS

To motivate the importance of clustering XML data, we summarize its use in several
data application domains.
2.1. XML Query Processing

Although XML is mainly perceived as a standard medium of information exchange,
storing, indexing, and querying of XML data are still important issues and have become research hotspots both in the academic and in the industrial communities [Wang
et al. 2004; Melton and Buxton 2006; Gou and Chirkova 2007]. Both XML native (e.g.,
Tamino, eXist, TIMBER) and enabled (e.g., Oracle, IBM DB2, SQL Server) Database
Management Systems (DBMSs) have been so far proposed [Bourret 2009] for storing and querying XML documents. Native DBMSs rely on a data model specifically
conceived for the management of XML, whereas enabled DBMSs are relational or
object-relational ones that have been extended for the treatment of XML. Enabled
XML DBMSs are more mature than the native ones because supported by big vendors
and the integration of XML data with other company data is easier. Some enabled
DBMSs support XML Schema for the specification of a mapping between an XSD and
internal relational or object-relational representation of XML documents [Florescu
and Kossmann 1999; Shanmugasundaram et al. 1999].
Query languages like Xpath and XQuery have been developed for accessing and manipulating XML documents in their native representation as well as extensions to the
SQL standard have been conceived to handle XML data besides relational data [Melton
and Buxton 2006]. All the standards so far proposed do not deal with data heterogeneity. To deal with heterogeneous queries, approximation techniques have been proposed
to evaluate twig patterns [Gou and Chirkova 2007]. Twig patterns are simple treestructured queries for XML that include three basic language elements, namely node
conditions, parent-child edges, and ancestor-descendant edges. Twig patterns are applicable to information-retrieval (IR) as well as database settings. Database-style queries
return all results that precisely match the content and structure requirements of the
query, while, IR-style queries allow fuzzy results, which are ranked based on their
query relevance.
Although twig pattern matching has become an important research area and several approaches have been developed to tackle it, it suffers from several drawbacks,
especially in large scale XML data and complex twig patterns, where data related to
the query appear in a small part of the whole XML document. So if we can access only
parts of the data that we need, the query processing can be conducted more efficiently
because the search space is reduced by skipping unnecessary data during the query
processing. A good solution is thus to consider clustering approaches in order to partition the whole XML data based on their common content, semantics, and structures
[Lian et al. 2004; Choi et al. 2007].
2.2. XML Data Integration

XML is widely used as the medium of data exchange among Web applications and
enterprises. Integration of distributed XML data is thus becoming a research problem.
This is due to the large amount of business data appearing on the Web; a large amount

25:6

A. Algergawy et al.

of service-oriented architecture is being adapted in the form of Web services. XML
data integration includes the construction of a global view for a set of independently
developed XML data [Batini et al. 1986; Le et al. 2006; Bertino and Ferrari 2001].
Since XML data are engineered by different people, they often have different structural and terminological heterogeneities. The integration of heterogeneous data sources
requires many tools for organizing and making their structure and content homogeneous. XML data integration is a complex activity that involves reconciliation at different levels: (1) at schema level, reconciling different representations of the same entity
or property, and (2) at instance level, determining if different objects coming from different sources represent the same real-world entity. Moreover, the integration of Web
data increases the integration process challenges in terms of heterogeneity of data.
Such data come from different resources and it is quite hard to identify the relationship with the business subjects. Therefore, a first step in integrating XML data is to
find clusters of the XML data that are similar in semantics and structure [Lee et al.
2002; Viyanon et al. 2008]. This allows system integrators to concentrate on XML data
within each cluster. We remark that reconciling similar XML data is an easier task
than reconciling XML data that are different in structures and semantics, since the
later involves more restructuring.
There are two directions of using clustering in XML data integration.
(1) A similarity matrix across XML data is determined and a clustering algorithm is
applied to the computed similarity matrix producing clusters of similar XML data.
Then, the XML data within each cluster are integrated [Lee et al. 2002].
(2) Each XML data tree is clustered into subtrees, which reduces the number of comparisons dramatically. The similarity degree based on data and structures of each
pair of subtrees is then measured. The data tree similarity degree is calculated
from the mean value of similarity degrees of matched subtrees. If the data tree
similarity degree is greater than a given threshold, the two XML documents can
be integrated [Viyanon et al. 2008].
2.3. XML Information Retrieval

Traditional Information Retrieval (IR) systems [Singhal 2001] rely either on the
Boolean model or the Vector Space model to represent the flat structure of documents as
a bag of words. Extensions of these models have been proposed, for example, the fuzzy
Boolean model and the knowledge-aware model. However, all of these indexing models
do ignore the structural organization of text. XML documents have a hierarchical structure defined by a DTD or an XML schema. While this structure allows documents to be
represented with hierarchical levels of granularity in order to achieve better precision
by means of focused retrieval, it implies more requirements on the representation and
retrieval mechanisms. The retrieval of XML documents using IR techniques is known
as XML-IR. Growing enthusiasm centering around XML retrieval led to the formation
of the Initiative for the Evaluation of XML Retrieval (or INEX in short).1 Organized
each year since 2002, INEX is a TREC2 -like forum where participating researchers can
discuss and evaluate their retrieval techniques using uniform scoring procedures over
a reasonably large relevance-assessed test collection. With the growing popularity of
XML clustering techniques, INEX 2009 includes a clustering track in which clustering
is used to organize a very large data set in a minimal number of clusters that need to
be searched to satisfy a given query and/or reorganize results furnished by an initial
search system as response to a user’s query.
1 http://inex.is.informatik.uni-duisburg.de/.
2 http://trec.nist.gov/.



25:7

An XML-IR process starts when a user submits a query into the system. Executing
the user query on the huge amount of XML documents is a time-consuming and errorprone process. A nice solution is to first cluster together semantically and structurally
similar XML documents. Then, the user query is executed on one or more related clusters [Lian et al. 2004; Tagarelli and Greco 2006]. Another way clustering can improve
the XML-IR process is using the clustering results for ranking the documents to be
returned to the user. The process is to take the search results of an XML search engine,
cluster them, and present them to the user in semantically distinct groups [Vutukuru
et al. 2002]. This assists in obtaining a unique cluster containing all relevant documents
or a set of clusters addressing the different aspects of the relevant information.
2.4. Web Mining

With the huge amount of information available online, the Web is a fertile area for
data mining research. Clustering XML data is a relevant problem in Web mining and
consists of the process of organizing data circulated over the Web into groups/clusters
in order to facilitate data availability and accessing, and at the same time to meet user
preferences. In an effort to keep up with the tremendous growth of the Web, many
approaches and systems have been proposed in order to organize their contents and
structures to make it easier for users to efficiently and accurately find the information they want. According to the type of mined information and the goal of the mining
process, these methods can be broadly classified into three categories. [Vakali et al.
2004; Pal et al. 2002]: (1) Web structure mining, referring broadly to the process of
uncovering interesting and potentially useful knowledge about the Web, (2) Web usage
mining using the Web-log data coming from users’ sessions to group together a set of
users’ navigation sessions having similar characteristics, and (3) Web content mining
clustering methods to try to identify inherent groupings of pages so that a set of clusters is produced in which relevant pages (to a specific topic) and irrelevant pages are
separated in the clustering process. XML clustering techniques can help in conceptualizing Web structure, usage, and content mining with a better use of structural and
content information represented in XML documents.
2.5. Bioinformatics

Bioinformatics represents a new field of scientific inquiry, devoted to answering questions about life and using computational resources to answer those questions. A key
goal of bioinformatics is to create database systems and software platforms capable of
storing and analyzing large sets of biological data. To that end, hundreds of biological
databases are now available and provide access to a diverse set of biological data. Given
the diversity and the exponential growth of biological data sets, and the desire to share
data for open scientific exchange, the bioinformatics community is continually exploring new options for data representation, storage, and exchange. In the past few years,
many in the bioinformatics community have turned to XML to address the pressing
needs associated with biological data [Cerami 2005].
XML-like presentations have been proposed for the following bio-molecular data
types.
(1) Principal bio-molecular entities (DNA, RNA, and proteins). For example, the Bioinformatic sequence markup language (BSML) [Hucka et al. 2003] has been used
to describe biological sequences (DNA, RNA, protein sequences), while ProXML
[Hanisch et al. 2002] is used to represent protein sequences.

25:8

A. Algergawy et al.

(2) Biological expression (microarray). The MAGE project3 provides a standard for the
representation of microarray expression data to facilitate their exchange among
different data systems.
(3) System biology. The need to capture the structure and content of biomolecular and
physiological systems has led to develop the System Biology Markup Language
(SBML).4
In general, clustering analysis can be used in two main directions: gene clustering [Andreopoulos et al. 2009; Tamayo et al. 1999; Eisen et al. 1998] and DNA or
protein sequence clustering [Sasson et al. 2002; Somervuo and Kohonen 2000]. Results
of gene clustering may suggest that genes in the same group have similar features
and functionalities, or they share the same transcriptional mechanism. The authors in
Jeong et al. [2006] propose a scheme for grouping and mining similar elements with
structural similarities from an XML schema for biological information, in which a number of elements and attributes are defined. cluML, a free, open, XML-based format, is a
new markup language for microarray data clustering and cluster validity assessment.
This format has been designed to address some of the limitations observed in traditional formats, such as inability to store multiple clustering (including biclustering)
and validation results within a data set [Bolshakova and Cunningham 2005]. On the
other hand, several clustering techniques have been proposed and applied to organize
DNA or protein sequence data. CONTOUR is a new approach that mines a subset of
high-quality subsequences directly in order to cluster the input sequences [Wang et al.
2009].
3. DATA REPRESENTATION

XML data can be represented using a common data model, such as rooted labeled trees,
directed acyclic graphs, or vector-based techniques. The data model should capture both
content and structure features of XML data and it is the basis for the identification
of the features to be exploited in the similarity computation (see Section 4). Data
representation starts with parsing XML data using an XML parsing tool, such as
the SAX parser.5 In the case of XML schema clustering, the parsing process may be
followed by a normalization process to simplify the schema structure according to a
series of predefined transformation procedures similar to those in Lee et al. [2002]. In
the remainder of the section the two most commonly used models to represent XML
data are discussed.
3.1. Tree-Based Representation

XML data can be represented as a data tree. A data tree (DT) is a rooted labeled tree
defined as a 3-tuple DT = (NT , ET , Lab NT ), where the following holds.
—NT = {nroot , n2 , . . . , nn} is a finite set of nodes, each of which is uniquely identified by
an object identifier (OID), where nroot is the tree root node. Three types of nodes in a
data tree can basically occur.
(1) Element nodes. They correspond to element declarations or complex type definitions in XML schemas or to tags in XML documents.
(2) Attribute nodes. These correspond to attribute declarations in XML schemas or
to attributes associated to tags in XML documents.
(3) Text nodes. These correspond to values in XML documents and basic types in
XML schemas.
3 MAGE

project: http://www.mged.org/Workgroups/MAGE/mage.html.

4 http://sbml.org/.
5 http://www.saxproject.org.



(a) Data tree of schema D1.

25:9

(b) Data tree of XML document.

Fig. 3. Tree representation of XML data.

—ET = {(ni , n j )|ni , n j ∈ NT } is a finite set of edges, where ni is the parent of n j . Each
edge represents a relationship between two nodes.
—Lab NT is a finite set of node labels. These labels are strings for describing the properties of the element and attribute nodes, such as name, data type, and cardinality,
or they are the data values associated with text nodes.
Figure 3 illustrates the tree representation of XML data. Specifically, Figure 3(a)
shows the data tree of schema D1 represented in Figure 1(a), while Figure 3(b) represents an XML document depicted in Figure 1(b). Each node in the data tree is associated
with the name label, (such as “Author” and “Title”) as well as its OID, (such as n1 and
n2 ). In Figure 3(a), the nodes n1 , n2 , and n4 represent examples of element nodes, and
node n6 is an attribute node. In Figure 3(b), node n9 is a text node. A data tree DT
is called an ordered labeled tree if a left-to-right order among siblings in DT is given,
otherwise it is called an unordered tree.
Given a tree representation of XML data, an object (OP) is a single portion to be
exploited for the computation of similarity. The property set associated to each object
is called the object feature. We classify objects in a data tree into: complex objects,
which include the whole data tree, subtrees, and paths; and simple objects, which
include element, attribute, and text nodes. Furthermore, there exist many relationships
among simple objects that reflect the hierarchical nature of the XML data tree. These
relationships include:
—parent-child (induced) relationships, that is the relationship between each element
node and its direct subelement/attribute/text node;
—ancestor-descendant (embedded) relationships, that is the relationship between each
element node and its direct or indirect subelement/attribute/text nodes;
—order relationships among siblings.
The tree-based representation of XML data introduces some limitations for the presentation of these relationships. For example, association relationships that are structural relationships specifying that two nodes are conceptually at the same level, are
not included in the tree representation. Association relationships essentially model
key/keyref and substitution group mechanisms. As a result, another tree-like structure,
such as directed acyclic graphs, is used [Boukottaya and Vanoirbeek 2005]. Figure 4(a)
represents a substitution group mechanism, which allows customer names to be in an
English or a German style. This type of relationship cannot be represented using the
tree representation. It needs a bidirectional edge, as shown in Figure 4(b).

25:10

A. Algergawy et al.

(a) Substitution group example.

(b) Data graph.

Fig. 4. Graph representation of an XML schema part.

3.2. Vector-Based Representation

The Vector Space Model (VSM) model [Salton et al. 1975] is a widely used data representation for text documents. In this model each document is represented as a feature
vector of the words that appear in documents of the data set. The term weights (usually term frequencies) of the words are also contained in each feature vector. VSM
represents a text document, docx , using the document feature vector, dx , as [Yang et al.
2009]:
dx = dx(1) , dx(2) , . . . , dx(n)

T

, dx(i) = TF(ρi , doxx ) · IDF(ρi ),

|D|
where TF(ρi , docx ) is the frequency of the term ρi of docx , IDF(ρi ) = log( DF(ρi ) ) is
the inverse document frequency of the term ρi for discounting the importance of the
frequently appearing terms, |D| is the total number of documents, DF(ρi ) is the number of documents containing ρi , and n is the number of distinct terms in the document set. Applying VSM directly to represent semistructured documents is not desirable, as the document syntactic structure tagged by their XML elements will be
ignored.
XML data generally can be represented as vectors in an abstract n-dimensional
feature space. A set of objects can be extracted from an XML data. Each object has a
set of features, where each feature is associated with a domain to deﬁne its allowed
values. The level of an XML element is a feature whose domain are the positive integers
(0 for the root, 1 for the ﬁrst level, and so on), while the name of the element is another
feature, whose domain is string. This representation model introduces a challenge
to incorporate the structural relationships between elements. In the following, we
elaborate on different methods used to model XML data based on their either content,
structure, or content and structure.

—Content-based representation. Each XML document, docx , is represented using a
vector, called the feature vector dx , that stores the frequency of distinct words in the
document. Figures 5(a) and (b) show the content-based representation of D1 depicted
in Figure 1(a).
—Structure-based representation. In this case, the feature vector captures the structure
information of the XML document instead of the content information. Several methods have been proposed to achieve this task. Most of them are based on exploiting
path information. The approach in Tran et al. [2008] models each XML document as a
vector {P1 , P2 , . . . , Pn}, where each element, Pi , of the vector represents the frequency
of the path that appears in the document, as shown in Figure 5(c). A path, Pi , contains element names from the root element to a leaf element. The leaf element is an
element that contains the textual content. The bitmap index technique is also used
to represent XML data [Yoon et al. 2001]. A set of XML documents is represented


25:11

(a) XML document instance of D1.
XML doc.
path
P1 (Book/Title)
P2 (Book/Author/Name)
P3 (Book/Publisher/Name)
P4 (Book/Author)
P5 (Book/Publisher)
P6 (Book/Year)
P7 (Conference/ConfTitle)
P8 (Conf/ConfAuthor)
P9 (Conf/ConfName)
P10 (Conf/ConfYear)
P11 (Conf/ConfLoc)

D1
1
1
1
0
0
0
0
0
0
0
0

D2
1
0
0
2
1
1
0
0
0
0
0

D3

(b) Feature vector (dx ) and feature matrix (Δx ) of D1.

D4

1
0
0
2
0
0
0
0
0
0
0

0
0
0
0
0
0
1
1
1
1
1

D5

D6

0
0
0
0
0
0
1
1
1
0
1

1
1
1
0
0
0
0
0
0
0
0

XML doc.
P1
P2
P3
P4
P5
P6
P7
P8
P9
P10
P11

(c) A structure-based representation for Fig.1(b).

D1

D2

D3

D4

D5

D6

1
1
1
0
0
0
0
0
0
0
0

path

1
0
0
1
1
1
0
0
0
0
0

1
0
0
1
0
0
0
0
0
0
0

0
0
0
0
0
0
1
1
1
1
1

0
0
0
0
0
0
1
1
1
0
1

1
1
1
0
0
0
0
0
0
0
0

(d) A bitmap index for Fig. 1(b).

Fig. 5. Vector representation of an XML document.

using a 2-dimensional matrix. As illustrated in Figure 5(d), if a document has path,
then the corresponding bit in the bitmap index is set to 1. Otherwise, all bits are set
to 0.
—Content and structure-based representation. Representing XML documents using either the content or the structure feature is not sufﬁcient to effectively model them;
both features should be taken into account. An extended version of the vector space
model called structural link vector model (SLVM) is used to capture syntactic structure tagged by XML elements [Yang et al. 2009]. SLVM represents an XML document
docx using a document feature matrix x ∈ Rn×m, given as:
x

=

x(1) ,

x(2) , . . . ,

x(m)

,

where m is the number of distinct XML elements, x(i) ∈ Rn is the TFIDF feature vector representing the ith XML element, (1 ≤ i ≤ m), given as x(i) =
T F(ρ j , docx .ei ).I DF(ρ j ) for all j = 1 to n, where T F(ρ j , docx .ei ) is the frequency of
the term w j in the element ei of docx . The SLVM representation of an XML document
instance for D1 depicted in Figure 1(a) is reported in Figure 5(b), which illustrates,
for example, that the term XML appears one time in D1 (from the document feature
vector dx ) under the element title (from the document feature matrix x ).
Another vector-based representation that captures both structure and content of
the XML data is represented in Yoon et al. [2001]. The bitmap indexing technique,
shown in Figure 5(d) is extended, where a set of XML documents is represented
using a 3-dimensional matrix, called BitCube. Each document is deﬁned as a set of
( path, word), where path is a root-to-leaf path, and word denotes the word or content
of the path. If a document has path, then the corresponding bit in the bitmap index
is set to 1. Otherwise, all bits are set to 0 (and if path contains a word, the bit is set
to 1, and 0 otherwise).

25:12

A. Algergawy et al.

Fig. 6. Similarity measures.

4. SIMILARITY MEASURES AND COMPUTATION

Starting from the representation model of objects and their features, the similarity
between XML data can be identiﬁed and determined by exploiting objects, objects’
features, and relationships among them. There are various aspects that allow the
description and categorization of XML data similarity measures, such as the kind
of methodology being used, the kind of XML data representation, and the planned
application domain [Tekli et al. 2009]. In the following, for homogeneity of presentation,
we survey several XML similarity measures based on the data representation used, as
shown in Figure 6.
4.1. Tree-Based Similarity Approaches

The computation of similarity among XML data represented as data trees depends on
the exploited objects on which similarity functions are to be applied. A function, Sim,
is a similarity measure that exploits the features of objects as well as the relationships
among them, in order to determine the proximity between objects. It is represented as
Sim(OP1 , OP2 ), and its value ranges between 0 and 1, when the measure is normalized.
The value 0 means strong dissimilarity between objects, while the value 1 means strong
similarity. Based on the objects used to compute the similarity among XML data trees,
we classify the similarity measures into: element-level measures and structure-level
measures.
4.1.1. Element-Level Measures. These measures, also known as schema matching-based
methods, consider simple objects’ details such as name, data type, as well as the relationships between objects. In element-level measures, the similarity between XML
data is based on the computed similarities among their elements. The similarity between two simple objects in two document trees OP1 ∈ DT1 and OP2 ∈ DT2 can be
determined using the following equation:

Sim(OP1 , OP2 ) = ws × SSim(OP1 , OP2 ) + wx × C Sim(OP1 , OP2 ),
where SSim(OP1 , OP2 ) represents the simple similarity measure between two objects exploiting their features, such as name, data type, and constraint, while
C Sim(OP1 , OP2 ) represents the complex similarity measure between them, exploiting
the relationships of each object, and ws and wx are weights to quantify the importance of each measure, (ws , wx ≥ 0, ws + wx = 1). These computed similarities are then
aggregated to determine the semantic similarities among paths and XML data trees
themselves [Rahm and Bernstein 2001; Kade and Heuser 2008].

type1
string
string
decimal
float
float
integer

type2
string
decimal
float
float
integer
short

25:13
T ypesim
1.0
0.2
0.8
1.0
0.8
0.8

(a) Type compatibility table.

*
+
?
none

*
1
0.9
0.7
0.7

+
0.9
1
0.7
0.7

?
0.7
0.7
1
0.8

none
0.7
0.7
0.8
1

(b) Cardinality constraint table.

Fig. 7. Type compatible and cardinality tables.

—Simple Similarity Measures. These measures determine the similarity between two
objects OP1 ∈ DT 1 and OP2 ∈ DT 2 using their features.
(1) Name Similarity. Object names can be semantically similar (e.g. person, people)
or syntactically similar (e.g. name, fname). Hence, both semantic and syntactic
measures are included to determine the degree of similarity between objects’
names. Semantic measures rely heavily on external or user-defined dictionaries
such as WordNet. In the syntactic measures, each object names is decomposed
into a set of tokens T1 and T2 using a customizable tokenizer using punctuation,
upper case, special symbols, and digits, for example LastName → {Last, Name}.
The name similarity between the two sets of name tokens T1 and T2 is determined
as the average best similarity of each token with a token in the other set. It is
computed as follows.
t1 ∈T1 [maxt2 ∈T2 sim(t1 , t2 )] +
t2 ∈T2 [maxt1 ∈T1 sim(t2 , t1 )]
.
Nsim(T1 , T2 ) =
|T 1| + |T 2|
To measure the string similarity between a pair of tokens, sim(t1 , t2 ), several string similarity functions, such as the edit distance and n-grams, can be
used [Cohen et al. 2003].
(2) Data Type Similarity. The Built-in XML data types hierarchy6 can be used in
order to compute data type similarity. Based on the XML schema data type
hierarchy, a data type compatibility table is built, similar to the one used in
Madhavan et al. [2001], as shown in Figure 7(a). The figure illustrates that
elements having the same data types or belonging to the same data type category
have the possibility to be similar and their type similarities (T ypesim) are high.
The type similarity between two objects using the type compatibility table is:
T ypesim(type1, type2) = T ypeT able (type1, type2).
(3) Constraint Similarity. Another feature of an object that makes a small contribution in determining the simple similarity is its cardinality constraint. The
authors of XClust [Lee et al. 2002] have defined a cardinality table for DTD
constraints, as shown in Figure 7(b). Also, the authors of PCXSS [Nayak and
Tran 2007] have adapted the cardinality constraint table for constraint matching of XSDs. The cardinality similarity between two objects using the constraint
cardinality table is:
Cardsim(card1, card2) = CardinalityT able (card1, card2).
(4) Content Similarity. To measure the similarity between text nodes of XML documents, the content information of these nodes should be captured. To this end,
a token-based similarity measure can be used. According to the comparison
made in Cohen et al. [2003], the TFIDF ranking performed best among several
token-based similarity measures. TFIDF (Term Frequency Inverse Document
6 http://www.w3.org/TR/xmlschema-2/.


25:14

A. Algergawy et al.

Fig. 8. The context of an object.

Frequency) is a statistical measure used to evaluate how important a word is to
a document or corpus. The importance increases proportionally to the number
of times a word appears in the document but is offset by the frequency of the
word in the corpus. It is used here to assess the similarity between text nodes. In
this case, the content of these nodes can be considered as multisets (or bags) of
words. Given the content of two text nodes represented as texts cont1 and cont2 ,
The TFIDF measure between them is given by [Cohen et al. 2003]:
Contsim (cont1 , cont2 ) =

V (w, cont1 ) · V (w, cont2 ),
w∈cont1 ∩cont2

where
V (w, cont1 ) =

V (w, cont1 )
w

V (w, cont1 )2

,

and
V (w, cont1 ) = log (TFw,cont1 + 1) · log (IDFw ),
where, TFw,cont1 is the frequency of the word w in cont1 , and IDFw is the inverse
of the fraction of names in the corpus that contain w.
These simple similarity measures are then aggregated using an aggregation function,
such as the weighted sum, to produce the simple similarity between two objects.
SSim (OP1 , OP2 ) = wn × Nsim(name1, name2)
+ wt × Typesim (type1, type2)
+ wc × Cardsim (card1, card2)
+ wco × Contsim (cont1, cont2),
where wn, wt , wc and wco are weights to quantify the importance of each similarity
measure, and wn + wt + wc + wco = 1. In XML schema, the value for wco is set to 0,
while in the XML document context wt = wc = 0. wn is assigned a larger value with
respect to wt and wc while comparing XML schema elements. In fact, the tuning of
weight values is a challenge and needs more attention [Lee et al. 2007].
—Complex Similarity Measures. These measures determine the similarity between two
objects OP1 ∈ DT 1 and OP2 ∈ DT 2 using their relationships, and exploiting the
computed simple similarities between objects. In general, these measures depend on
the object (node) context, which is reflected by its ancestors and its descendants, as
shown in Figure 8. The descendants of an object include both its immediate children
and the leaves of the subtrees rooted at the element. The immediate children reflect
its basic structure, while the leaves reflect the element’s content. As a sequence,
these measures may depend on the following.


25:15

(1) Child Context. The child context of an object (node) is the set of its immediate
children. To compute the child context similarity between two objects OP1 and
OP2 , the child context set is first extracted for each node, then the simple similarity between each pair of children in the two sets is determined, the matching
pairs with maximum similarity values are selected, and finally the average of
best similarity values is computed.
(2) Leaf Context. The leaf context of an object (node) is the set of leaf nodes of subtrees
rooted at the node. To determine the leaf context similarity between two objects
OP1 and OP2 , the leaf context set is extracted for each node, then a suitable
comparison function between two sets is applied.
(3) Ancestor Context. The ancestor context of an object (node) is the path extending
from the root node to the node itself. To measure the ancestor context similarity
between the two objects OP1 and OP2 , each ancestor context , say path P1 for
OP1 and P2 for OP2 , has to be extracted. Then, the two paths are compared using
a suitable path comparison function.
These similarity measures are then aggregated using an aggregation function, such
as the weighted sum, to produce the complex similarity between two objects.
C Sim(OP1 , OP2 ) = CombineC (Child (OP1 , OP2 ),
Leaf (OP1 , OP2 ),
Ancestor (OP1 , OP2 )),
where Child (OP1 , OP2 ), Leaf(OP1 , OP2 ), and Ancestor (OP1 , OP2 ) are the similarity
functions that compute the child, leaf, and ancestor contexts between the two objects,
respectively, and CombineC is the aggregation function to combine the similarity values.
The schema matching community, in general, classifies approaches for XML schema
matching (element-level) into two main approaches [Rahm and Bernstein 2001], as
shown in Figure 6.
—Schema-Based Approaches. These approaches only consider schema information, not
instance data. Schema-based approaches exploit the object features and their relationships. These properties are then exploited using either individual matchers or
combining matchers in a rule-based fashion to determine semantic similarity among
schema elements. An individual matcher exploits only one type of element properties in a single matcher, while a combining matcher can be one of two types: hybrid
matchers, which exploit multiple types of element properties in a single algorithm
and composite matchers, which combine the results of independently executed matchers. The element level similarities are then aggregated to compute the total similarity
between schemas [Do and Rahm 2002; Melnik et al. 2002; Giunchiglia et al. 2007;
Bonifati et al. 2008; Saleem et al. 2008; Algergawy et al. 2009]. They are easy to
implement and do not need to be trained before being used.
—Instance-Based Approaches. These approaches consider data instances as well as
schema-based information. The instance-level data give important insight into the
contents and meaning of schema elements. The main advantage of these approaches
is that they can empirically learn the similarities among data, relying on their
instance values. Hence, many learner-based schema matching systems have been
developed to determine element-level similarity [Li and Clifton 2000; Doan et al.
2004]. These systems depend largely on prematch phases such as the training
phase using unsupervised learning in SemInt [Li and Clifton 2000], using machine
learning in GLUE [Doan et al. 2004], or using neural network-based partial
least squares in Jeong et al. [2008]. However, the main disadvantage of using
learner-based approaches is that instance data is generally available in very vast

25:16

A. Algergawy et al.

quantity. Hence, the computational cost is very expensive, which affects the schema
matching performance.
4.1.2. Structure-Level Measures. On the other hand, structure-level measures, also
known as tree-editing methods, exploit complex objects without taking into account
the detailed object components in the data tree. The tree-editing problem is the generalization of the problem of computing the distance between two strings to labeled
trees. As usual, the edit distance relies on three elementary edit operations: insertion,
deletion, and relabeling of a node.
Let DT1 and DT2 be two data trees and assume that we are given a cost function
defined on each edit operation. An edit script (ES) between the two data trees DT1 and
DT2 is a sequence of edit operations turning DT1 into DT2 . The cost of ES is the sum
of the costs of the operations in ES. An optimal edit script between DT1 and DT2 is
an edit script between DT1 and DT2 of minimum cost. This cost is called the tree edit
distance, denoted by δ(DT1 , DT2 ) [Bille 2005], that is:

δ(DT1 , DT2 ) = min{γ (ES)| ES is an edit operation sequence transforming DT1 to DT2 },
where γ is a cost function defined on each edit operation.
Several algorithms have been proposed to solve the tree edit distance problem. The
first nonexponential algorithm that allows the insertion, deletion, and relabeling of inner and leaf nodes has a complexity of O(|DT1 |×|DT2 |×depth(DT1 )2 ×depth(DT2 )2 ) [Tai
1979], where |DT1 | and |DT2 | are data tree cardinalities. Another set of approaches
has been proposed allowing the edit operations of nodes anywhere in the tree [Zhang
and Shasha 1989; Shasha and Zhang 1995]. These early attempts to solve the tree
edit distance problem were not mainly developed in the context of XML data similarity, and thus might yield results that are not completely adequate for XML data. The
work proposed in Chawathe [1999] has been considered as the keystone for various
XML-related structural comparison approaches [Tekli et al. 2009]. Chawathe suggests
a recursive algorithm to calculate the tree edit distance between two rooted ordered
labeled trees, using a shortest path detection technique on an edit graph. The author
restricts the insertion and deletion operations to leaf nodes, and allows the relabeling
of nodes anywhere in the data tree. Chawathe’s algorithm has an overall complexity of
O(|DT1 | × |DT2 |).
Even if Chawathe’s algorithm is considered as a starting point for recent XML tree
edit distance approaches, the algorithm lacks subtree similarity computation. Due
to the frequent presence of repeated and optional elements in XML data (especially
XML documents), there is a growing need to identify subtree structural similarities
in the XML data tree comparison context. As a result, several approaches have been
proposed to address the subtree structural similarity [Nierman and Jagadish 2002;
Dalamagas et al. 2006]. The work provided in Nierman and Jagadish [2002] extends
Chawathe’s approach by adding two new operations, insert tree and delete tree, to
discover subtree similarities. While this algorithm outperforms, in quality, Chawathe’s
algorithm, the authors in Nierman and Jagadish [2002] show that their algorithm is
more complex than its predecessor, requiring a precomputation phase for determining
the costs of tree insert and delete operations. The algorithm provided in Dalamagas
et al. [2006] proposes using tree structural summaries that have minimal processing
requirements instead of the original trees representing the XML documents. Those
summaries maintain the structural relationships between the elements of an XML
document, reducing repetition and nesting of data tree elements. The authors present
a new algorithm to calculate tree edit distances and define a structural distance metric
to estimate the structural similarity between the structural summaries of two rooted
ordered labeled trees. The algorithm has a complexity of O(|DT1 | × |DT2 |).


25:17

Table I. Tree Edit Distance Algorithms
Method
[Tai 1979]
[Zhang and
Shasha 1989]

Edit operations
insert, delete,
relabel
insert, delete,
relabel

[Shasha and
insert, delete,
Zhang 1995]
relabel
[Chawathe 1999] insert, delete
relabel

[Nierman and
insert, delete
Jagadish 2002] relabel
insert tree
delete tree
[Dalamagas
insert, delete
et al. 2006]
relabel

Where
Time complexity
anywhere in the
O(|DT1 | × |DT2 |×
data tree
depth(DT1 )2 × depth(DT2 )2 )
anywhere in
O(|DT1 | × |DT2 |)×
the data tree
min(depth(DT1 ), leaf (DT1 ))×
min(depth(DT2 ), leaf (DT2 ))
anywhere in
O(|DT1 | × |DT2 |)×
the data tree
depth(DT1 ) × depth(DT2 )
leaf nodes
O(|DT1 | × |DT2 |)
anywhere in the
data tree
leaf nodes
anywhere in
the data tree
leaf nodes
anywhere in the
data tree

Comparison context
similarity between
ordered labeled trees
similarity between

similarity between
hierarchical
structured data such
as object class
hierarchies, HTML,
XML
O(2 × |DT | + |DT1 | × |DT2 |) Structural similarity
in XML documents
O(|DT1 | × |DT2 |)

Structural similarity
in XML document
clustering

Fig. 9. Tree distance between XML data.

Table I reports the tree edit distance algorithms representing their time complexity
and their comparison contexts. The table shows that the last three methods ﬁt for measuring the similarity between XML data. However, they only consider the structure of
XML data and ignore their semantics and content. Figure 9 illustrates that the edit
operations required to transform DT 1 to DT 2 are equal to that required to transform
DT 2 to DT 3 because only one relabeling operation is required in both cases to transform the source tree into the target tree. A dotted line from a node in a data tree, such
as DT 1, to a node in another data tree, such as DT 2, indicates that a relabeling operation is required. Assigning a constant cost for the edit operations results in equal tree
distances between DT 1 and DT 2 and DT 2 and DT 3. This simple example shows that
the tree editing method may not be able to distinguish the hierarchical difference in
some situations and a combination of simple and complex similarity measures should
be applied.
A number of approaches to measure the similarity between XML data have been
proposed. These approaches can be classiﬁed into three types [Nayak and Iryadi 2007],
as shown in Figure 6.
—Measuring the structure similarity between XML documents. To measure the structure similarity of heterogenous XML documents, XML element/attribute values are
generally ignored. Different types of algorithms have been proposed to measure the

25:18

A. Algergawy et al.

document similarity based on their structure, including tree-edit distance similarity, path similarity, set similarity, and so on [Buttler 2004; Tekli et al. 2009]. The
authors in Buttler [2004] and Rafiei et al. [2006] represent the structure of XML
documents as a set of paths from the root to a leaf. They also consider any partial
path, the path from the root to any node of the data tree. XML documents are then
compared according to their corresponding sets of paths. The experimental results
in Buttler [2004] and Rafiei et al. [2006] show that the path similarity method provides accurate similarity results compared with the tree-edit distance results. While,
the authors in Candillier et al. [2005] transform data trees into sets of attributevalues, including the set of parent-child relations, the set of next-sibling relations,
and the set of distinct paths. The authors can then apply various existing methods
of classification and clustering on such data, using the structural description of XML
documents alone. However, the set similarity method is not compared to similarly
existing methods, such as the tree-edit similarity or the path similarity.
—Measuring the structure and content similarity between XML documents. While several methods have been proposed to measure the similarity between XML documents
based on their structural features, others consider the content feature in their similarity computation. This has been an active area of research and it is fundamental to many applications, such as document clustering, document change detection,
integrating XML data sources, XML dissemination, and approximate querying of
documents [Leung et al. 2005]. A number of approaches have been used to measure
the document similarity considering both structure and content features of XML documents, such as leaf node clustering [Liang and Yokota 2005; Viyanon et al. 2008],
Baysian networks [Leito et al. 2007], pattern matching [Dorneles et al. 2004], and
so on. The authors in Liang and Yokota [2005] propose LAX (leaf-clustering-based
approximate XML join algorithm), in which two XML document trees are clustered
into subtrees representing independent items and the similarity between them is
determined by calculating the similarity degree based on the leaf nodes of each pair
of subtrees. The experimental results in Liang and Yokota [2005] illustrate that LAX
is more efficient in performance and more effective for measuring the approximate
similarity between XML documents than the tree edit distance. The authors state
that when applying LAX to large XML documents, the hit subtrees selected from the
output pair of fragment documents that have large tree similarity degrees might not
be the proper subtrees to be integrated. SLAX [Liang and Yokota 2006] is therefore
an improved LAX to solve this problem. To address the drawbacks of LAX and SLAX,
the authors in Viyanon et al. [2008] develop a three-phase approach. Each data tree
is first clustered into a set of subtrees by considering leaf-node parents as clustering
points. The clustered subtrees are then considered independent items that will be
matched. The best matched subtrees are integrated in the first XML tree as a resulting XML tree. The experimental results in Viyanon et al. [2008] show that their
developed approach performs better than LAX and SLAX.
—Measuring the structural similarity between data and schema. Evaluating similarity
between XML documents and their schemas can be exploited in various domains,
such as for classifying XML documents against a set of DTDs/schemas declared
in an XML data, XML document retrieval via structural queries, as well as the
selective dissemination of XML documents wherein user profiles are expressed as
DTDs/schemas, against which the incoming XML data stream is matched [Bertino
et al. 2004, 2008; Tekli et al. 2007]. Measuring the structural similarity between XML
documents and DTDs/XSDs is not a trivial task. Many factors should be taken into
account in the evaluation, like the hierarchical and complex structure of XML documents and schemas, and the tags used to label their semantics. These factors limit
developing approaches to measures of this kind of structural similarity [Bertino et al.


25:19

Fig. 10. XML document instance of D2 and its feature vector.

2008; Tekli et al. 2007]. The similarity approach in Tekli et al. [2007] relies on transforming both XML documents and DTDs into ordered label trees and then applying
the tree editing distance. The approach in Bertino et al. [2008] also represents XML
documents and DTDs as trees, and then computes the structural similarity between
a tree and an intensional representation of a set of trees (generated from the DTD).
This result could be achieved through an extensional approach, that is, by making
the set of document structures described by the DTD, explicit, and computing the
minimal distance between the document and each document structure. However,
this approach is infeasible because of the large number of document structures that
can be generated from a DTD.
4.2. Vector-Based Similarity Approaches

The similarity between XML data represented using vector-based approaches is mainly
based on the exploited features. As shown in Figure 6, the vector-based similarity approaches can exploit either content, structure, or both, of compared objects. In general,
once the features have been selected, the next step is to define functions to compare them. Given a domain Di , a comparison criterion for values in Di is defined as
a function Ci : Di × Di −→ Gi , where Gi is a totally ordered set, typically the real
numbers. A similarity function Sim : (D1 , . . . , Dn) × (D1 , . . . , Dn) −→ L, where L is
a totally ordered set, can now be defined to compare two objects represented as feature vectors and returns a value that corresponds to their similarity [Guerrini et al.
2007]. If feature vectors are real vectors, metric distances induced by norms are typically used. The best-known examples are the L1 (Manhattan) and L2 (Euclidean)
distances. Other measures have been proposed based on the geometric and probabilistic models. The most popular geometric approach to distance is the vector space
model used in Information Retrieval [Salton et al. 1975]. Other popular similarity
v2
v
measures are the cosine (cos(v1 , v2 ) = |vv11||v2 | ), Dice (Dice(v1 , v2 ) = |v2v21|v2 |2 ), and Jaccard
1|
2
v1 v2
(Jac(v1 , v2 ) = |v1 |2 |v2 |2 −v1 v2 ), coefficients. In the following, we elaborate on vector-based
similarity approaches guided by the taxonomy shown in Figure 6.
—Content-based similarity measures. In this case, each XML document is represented
as a feature vector exploiting the content feature of the document elements. The
similarity between XML document pairs can be determined using IR measures. For
example, to compute the similarity between two document instances of D1 and D2
(see Figure 5 and Figure 10), where dx is the feature vector of D1, and dy is the
feature vector of D2, the cosine measure can be used as given [Yang et al. 2009],
d .d
sim(D1, D2) = cos(dx , dy ) = |dxx||dyy | = 0.9258. Although the similarity value is high,
the two document instances, as stated previously, should be in separate clusters. This
conflict arises from the vector representation of document instances, which considers
only the content feature and ignores the structural features of the documents. To
consider the semantic associated to document contents, the approach in Tran et al.

25:20

A. Algergawy et al.

[2008] makes use of the latent semantic kernel [Cristianini et al. 2002]. Given two
XML documents, docx and doc y , represented as vectors, dx and dy , the semantic
similarity of the documents content is measured as:
Sim(docx , doc y ) = contSim (dx , dy ) =

T
dx PP T dy
,
T d ||P T d |
|P x
y

where P is a matrix used as a mapping function to transform the two documents,
dx and dy , into concept space to determine the semantic association of document
contents.
—Structure-based similarity measures. To measure the similarity between XML data
based on their structures requires exploiting the structure information of XML data
in the feature vectors. The xOR is used to determine the distance between two XML
documents represented in the BitCube [Yoon et al. 2001]. Given two XML documents
(two bitmap rows), docx and doc y , the similarity is given by:
Sim(docx , doc y ) = 1 −

|xOR(docx , doc y )|
,
max(|docx ||doc y |)

where xOR is a bit-wise exclusive OR operator, and |docx | and |doc y | are the cardinalities of the documents.
—Structure and content-based similarity measures. To measure the similarity between
XML documents exploiting both the content and structure features, two strategies
can be utilized.
(1) The ﬁrst strategy relies on representing the content and structure of XML documents using feature vectors. The content similarity and the structure similarity
are then computed separately, and the two computed values are combined using
an aggregation function [Tran et al. 2008].
(2) The second strategy is based using the representation of XML documents as
feature matrices that capture both the content and structure information of
them. A suitable similarity measure can be proposed. For example, to compute
the similarity between two XML documents represented in an n-dimensional
matrix is presented by introducing the kernel matrix:
n

Sim (docx , doc y ) =

T
dxi • Me • dyi ,
i=1

where dx and dy are the normalized document feature vectors of docx and doc y ,
respectively, Me is a kernel matrix that captures the similarity between pairs of
matrix elements, and • indicates the vector dot product.
Experimental evaluation reported in Tran et al. [2008] and Yang et al. [2009] shows
that combining the structure and content features to compute the similarity between
XML documents outperforms the other two methods, especially when documents belong
to different structural deﬁnitions.
5. CLUSTERING/GROUPING

XML data that are similar in structures and semantics are grouped together to form
a cluster using a suitable clustering algorithm [Jain et al. 1999; Berkhin 2002; Xu and
Wunsch 2005]. Clustering methods are generally divided into two broad categories:
hierarchical and nonhierarchical (partitional) methods. The nonhierarchical methods
group a data set into a number of clusters using a pairwise distance matrix that records
the similarity between each pair of documents in the data set, while the hierarchical
methods produce nested sets of data (hierarchies), in which pairs of elements or


25:21

clusters are successively linked until every element in the data set becomes connected.
Many other clustering techniques have been developed considering the more frequent
requirement of tackling large-scale and high-dimensional data sets in many recent
applications.
5.1. Hierarchical Methods

Hierarchical clustering builds a hierarchy, known as a dendrogram. The root node of
the dendrogram represents the whole data set and each leaf node represents a data
item. This representation gives a quite informative description and visualization of the
data clustering structure. Strategies for hierarchical clustering generally fall into two
types: agglomerative and division. The divisive clustering algorithm builds the hierarchical solution from the top toward the bottom by using a repeated cluster bisectioning
approach. Given a set of N XML data to be clustered, and an N × N similarity matrix,
the basic process of the agglomerative hierarchical clustering algorithm is as follows.
(1) Treat each XML data of the data set to be clustered as a cluster, so that if you have
N items, you now have N clusters, each containing just one XML data.
(2) Find the most similar pair of clusters and merge them into a single cluster.
(3) Compute similarities between the new cluster and each of the old clusters.
(4) Repeat steps 2 and 3 until all items are clustered into a single cluster of size N.
Based on different methods used to compute the similarity between a pair of clusters
(Step 3), there are many agglomerative clustering algorithms. Among them are the
single-link and complete-link algorithms [Jain et al. 1999; Manning et al. 2008]. In
the single-link method, the similarity between one cluster and another is equal to the
maximum similarity from any member of one cluster to any member of the other cluster.
In the complete-link method, the similarity between one cluster and another has to be
equal to the minimum similarity from any member of one cluster to any member
of the other cluster. Even if the single-link algorithm has a better time complexity
(O(N 2 )) than the time complexity (O(N 2 log N)) of the complete-link algorithm, it has
been observed that the complete-link algorithm produces more useful hierarchies in
many applications than the single-link algorithm. Further, the high cost for most of
the hierarchical algorithms limits their application in large-scale data sets. With the
requirement for handling large-scale data sets, several new hierarchical algorithms
have been proposed to improve the clustering performance, such as BIRCH [Zhang
et al. 1996], CURE [Guha et al. 1998], and ROCK [Guha et al. 2000].
5.2. Nonhierarchical Methods

The nonhierarchical clustering methods produce a single partition of the data set
instead of a clustering structure, such as the dendrogram produced by a hierarchical
technique. There are a number of such techniques, but two of the most prominent
are k-means and K-medoid. The nonhierarchical clustering methods also have similar
steps as follows.
(1) Select an initial partition of the data with a ﬁxed number of clusters and cluster
centers;
(2) Assign each XML data in the data set to its closest cluster center and compute the
new cluster centers. Repeat this step until a predeﬁned number of iterations or
there is no reassignment of XML data to new clusters; and
(3) Merge and split clusters based on some heuristic information.
The K-means algorithm is popular because it is easy to implement, and it has a
time complexity of O(NKd), where N is the number of data items, K is the number of
clusters, and d is the number of iterations taken by the algorithm to converge. Since K

25:22

A. Algergawy et al.

and d are usually much less than N and they are fixed in advance, the algorithm can
be used to cluster large data sets. Several drawbacks of the K-means algorithm have
been identified and well studied, such as the sensitivity to the selection of the initial
partition, and the sensitivity to outliers and noise. As a result, many variants have
appeared to overcome these drawbacks [Huang 1998; Ordonez and Omiecinski 2004].
5.3. Other Clustering Methods

Despite their widespread use, the performance of both hierarchal and nonhierarchal
clustering solutions decreases radically when they are used to cluster a large scale
and/or a large number of XML data. This becomes more crucial when dealing with
large XML data as an XML data object is composed of many elements and each
element of the object individually needs to be compared with elements of another
object. Therefore the need for developing another set of clustering algorithms arises.
Among them are the incremental clustering algorithms [Can 1993; Charikar et al.
2004] and the constrained agglomerative algorithms [Zhao and Karypis 2002b].
Incremental clustering is based on the assumption that it is possible to consider XML
data one at a time and assign them to existing clusters with the following steps.
(1) Assign the first XML data to a cluster.
(2) Consider the next XML data object and either assign it to one of the existing clusters
or assign it to a new one. This assignment is based on the similarity between the
object and the existing clusters.
(3) Repeat step 2 until all the XML data are assigned to specific clusters.
The incremental clustering algorithms are noniterative—they do not require computing the similarity between each pair of objects. So, their time and space requirements
are small. As a result, they allow clustering large sets of XML data efficiently, but
they may produce a lower quality solution because they avoid computing the similarity
between every pair of documents.
Constrained agglomerative clustering is a trade-off between hierarchical and nonhierarchical algorithms. It exploits features from both techniques. The constrained
agglomerative algorithm is based on using a partitional clustering algorithm to constrain the space over which agglomeration decisions are made, so that each XML data
is only allowed to merge with other data that are part of the same partitionally discovered cluster [Zhao and Karypis 2002b]. A partitional clustering algorithm is first
used to compute a k-way clustering solution. Then, each of these clusters, referred to as
constraint clusters, is treated as a separate collection, and an agglomerative algorithm
is used to build a tree for each one of them. Finally, the k different trees are combined
into a single tree by merging them using an agglomerative algorithm that treats the
documents of each subtree as a cluster that has already been formed during agglomeration. The constrained agglomerative algorithms has two main advantages. First, it is
able to benefit from the global view of the collection used by partitional algorithms and
the local view used by agglomerative algorithms. An additional advantage is that the
computational complexity of the algorithm is O(k( N )2 log( N )), where√ is the number
k
k
k
of constraint clusters. If k is reasonably large, for example, k equals N, the original
2
complexity of O(N 2 log N) for agglomerative algorithms is reduced to O(N 3 log N).
5.4. Data Clustering Evaluation Criteria

The performance of XML data clustering approaches can be evaluated using two major
aspects: the quality and the efficiency of clusters.
5.4.1. Quality Evaluation. The quality of clusters can be measured by external and internal criteria [Guerrini et al. 2007]. The external quality measures, such as the entropy,


25:23

purity, and FScore, use an external manual classification of the documents, whereas
the internal quality measures are evaluated by calculating the average inter- and
intra-clustering similarity.
FScore is a trade-off between two popular information retrieval metrics, precision
P and recall R [Baeza-Yates and Ribeiro-Neto 1999]. Precision considers the rate of
correct matches in the generated solution, while recall considers the rate of correct
matches in the model solution. Given a cluster Ci , let TP be the number of XML data
in Ci that are similar (correctly clustered), FP be the number of documents Ci that
are not similar (misclustered), FN be the number of documents that are not in Ci but
should be, N be the total number of XML data, and Ni be the number of XML data in
Ci . The precision Pi and recall Ri , of a cluster Ci , are defined as follows.
Pi =

TP
TP
, Ri =
.
TP + FP
TP + FN

(1)

FScore combining precision and recall with equal weights for the given cluster Ci is
defined as:
Pi × Ri
F Scorei = 2 ×
.
(2)
Pi + Ri
Hence, the FScore of the overall clustering approach is defined as the sum of the
individual class FScores weighted differently according to the number of XML data in
the class:
k
i=1

Ni × F Scorei
,
(3)
N
where k is the number of clusters. A good clustering solution has the FScore value
closer to one. A number of methods use two standard measures derived from FScore to
reflect the quality of each cluster and intercluster: micro F1 and macro F1 measures.
Micro-average FScore is calculated by summing up the TP, the FP, and the FN values
from all the categories; FScore value is then calculated based on these values. Macroaverage FScore, on the other hand, is derived from averaging the FScore values over
all the categories. The best clustering solution for an input data set is the one where
micro- and macro-average FScore measures are close to 1.
Purity, in general, is a quantitative assessment of homogeneity. Hence, purity measures the degree to which a cluster contains XML data primarily from one class. The
purity of cluster Ci is defined as [Zhao and Karypis 2002a]:
F Score =

Pur(Ci ) =

1
max Nir ,
Ni

(4)

which is nothing more than the fraction of the overall cluster size (Ni ) that represents
the largest class of documents (Nir ) assigned to that cluster. The overall purity of the
clustering solution is obtained as a weighted sum of the individual cluster purities and
it is:
k

purity =
i=1

Ni
Pur(Ci ).
N

(5)

In general, the larger the values of purity, the better the clustering solution is.
Entropy is a widely used measure for clustering solution quality, which measures how
the various classes of the XML data are distributed within each cluster. The entropy of

25:24

A. Algergawy et al.

a cluster Ci is defined as:
E(Ci ) = −

q

1
log q

log
r=1

Nir
,
Ni

(6)

where q is the number of classes in the XML dataset, and Nir is the number of XML data
of the rth class that is assigned to the cluster ith. The entropy of the entire clustering
solution is then defined to be the sum of the individual cluster entropies weighted
according to the cluster size. That is,
k

Entropy =
i=1

Ni
E(Ci ).
N

(7)

A perfect clustering solution will be the one that leads to clusters that contain documents from only a single class, in which case the entropy will be zero. In general, the
smaller the entropy values, the better the clustering solution is.
The internal clustering solution quality is evaluated by calculating the average interand intra-clustering similarity. The intra-clustering similarity measures the cohesion
within a cluster, how similar the XML data within a cluster are. This is computed by
measuring the similarity between each pair of items within a cluster, and the intraclustering similarity of a clustering solution is determined by averaging all computed
similarities taking into account the number of XML data within each cluster:
IntraSim =

k
i=1

IntraSim(Ci )
.
N

(8)

In general, the larger the values of intra-clustering similarity (IntraSim), the better
the clustering solution is.
The inter-clustering similarity measures the separation among different clusters.
It is computed by measuring the similarity between two clusters. A good clustering
solution has lower inter-clustering similarity values.
5.4.2. Efficiency Evaluation. Efficiency (scalability) is one of the most important aspects
of today’s software applications. As XML data grow rapidly, the systems that support
XML data management need to grow. As they grow, it is important to maintain their
performance (both quality and efficiency). The scalability of a system, such as XML data
clustering, is a measure of its ability to cost-effectively provide increased throughput,
reduced response time, and/or support more users when hardware resources are added.
From this definition, efficiency (scalability) is mainly evaluated using two properties:
speed (the time it takes for an operation to complete), and space (the memory or nonvolatile storage used up by the construct). In order to obtain a good efficiency measure
for XML data clustering approaches, we should consider two factors [Algergawy et al.
2008a]. The first factor is the identification of the critical phase of a clustering process,
data representation, similarity computation, or grouping/clustering. Intuitively, data
representation is a basic phase, while both similarity computing and grouping are two
critical phases that affect the XML data clustering efficiency. This leads to the second
factor, the type of methodology used to perform the required task. Is the similarity computation performed pairwise or holistically? Is the clustering algorithm hierarchical,
nonhierarchical, or incremental?
6. XML DATA CLUSTERING PROTOTYPES: A COMPARISON

In this section we present current approaches for clustering XML data. The description relies on the kind of data to be clustered, documents or schemas, focusing on the


25:25

(a) XML document

(b) Level Structure

(c) Level Edge

Fig. 11. Example of Level Structure and LevelEdge.

adopted data representation among those discussed in Section 3, the employed similarity measure among those discussed in Section 4, and the clustering algorithm and
quality evaluation criteria among those discussed in Section 5.
6.1. Clustering XML Document Prototypes

6.1.1. XCLS. XML documents clustering with level similarity (XCLS) [Nayak 2008]
represents an incremental clustering algorithm that groups XML documents according
to their structural similarity. Following our generic framework, XCLS has the following
phases.

—Data Representation. XML documents are first represented as ordered data trees
where leaf nodes represent content values. XCLS considers only document elements
not document attributes. Each node (simple object) is ranked by a distinct integer
according to the preorder traversal of the data (document) tree. A level structure
format is introduced to represent the structure of XML documents for efficient processing, as shown in Figure 11(b). The level structure groups the distinct XML nodes,
distinguished by a number associated to each node as shown in Figure 11(a), for each
level in the document.
—Similarity Computation. XCLS proposes a new algorithm for measuring the structural similarity between XML documents represented as level structures, called level
similarity. Level similarity measures the occurrences of common elements in each
corresponding level. Different weights are associated to elements in different levels.
Consider two level representations L1 and L2 , with N1 and N2 levels, respectively,
and a positive integer a > 0. The level similarity, SimL1 ,L2 , between two level representations is:
SimL1 ,L2 =

0.5 ×

M−1 1
i=0 ci

× aM−i−1 + 0.5 ×
M−1
k=0 tk

M−1 2
j=0 c j

× aM−k−1

× aM− j−1

,

1
where ci denotes the number of common elements in level i of L1 and some levels of
L2 , where c2 denotes the number of common elements in level j of L2 and some levels
j
of L1 during the matching process, tk denotes the total number of distinct elements in
the level k of L1 , and M is the number of levels in the first level structure. Since the
level similarity measure is not transitive, XCLS determines the structural similarity
from the first XML document to the second one and vice versa and chooses the highest
value between the two.
—Clustering/Grouping. In this phase, an incremental clustering algorithm is used
to group the XML documents within various XML sources, considering the level
similarity. The clustering algorithm progressively places each coming XML document
into a new cluster or into an existing cluster that has the maximum level similarity
with it.


25:26

A. Algergawy et al.

—Evaluation Criteria. The quality of XCLS is evaluated using the standard criteria,
such as intra- and inter-cluster similarity, purity, entropy, and FScore. The scalability
of the XCLS system is also evaluated to validate its space and time complexity. Since
XSLC uses an incremental clustering algorithm that avoids the need of pairwise
comparison, it has a time complexity of O(NKdc), where c is the number of distinct
elements in clusters. Experimental evaluation indicates that XCLS achieves a similar
quality result as the pairwise clustering algorithms in much less time. However,
XCLS lacks a unified clustering framework that can be applied efficiently both to
homogeneous and heterogenous collections of XML documents.
6.1.2. XEdge. XML documents clustering using edge level summaries (XEdge)
[Antonellis et al. 2008] represents a unified clustering algorithm for both homogeneous
and heterogenous XML documents. Based on the type of the XML documents, the
proposed method modifies its distance metric in order to adapt to the special structure
features of homogeneous and heterogeneous documents.

—Data Representation. Like XCLS, XEdge represents XML documents as ordered data
trees where leaf nodes represent content values. Instead of summarizing the distinct
nodes as XCLS, XEdge introduces the LevelEdge representation, which summarizes
all distinct parent/child relationships, distinguished by a number associated to each
edge as shown in Figure 11(a), in each level of the XML document, as shown in
Figure 11(c). The distinct edges are first encoded as integers and those integers are
used to construct the LevelEdge representation.
—Similarity Computation. In order to compute the similarity among XML documents,
XEdge proposes two similarity measures to distinguish between similarity computation among homogeneous XML documents and similarity computation among
heterogeneous ones. Assuming that homogeneous XML documents are derived from
subDTDs of the same DTD, XEdge only searches for common edges in the same levels
in both documents. To measure the similarity between two homogeneous XML documents represented as LevelEdge L1 and L2 , it uses the following similarity function.
SimL1 ,L2 =

m−1
i=0 ci
M−1
k=0 tk

× am−i−1
× aM−k−1

,

where ci is the number of common distinct edges in the level i of L1 and L2 , while tk is
the total number of distinct edges in the level k of both L1 and L2 , m = min(N1 , N2 ),
and M = max(N1 , N2 ). For heterogeneous XML documents, common edges in different levels should be identified and a similarity function similar to the one used in
XCLS is used.
—Clustering/Grouping. In this phase, a partitional clustering algorithm based on a
modified version of k-means is used.
—Evaluation Criteria. The quality of the XEdge system is only evaluated using external
criteria, such as precision, recall, and FScore. The experimental results show that
both XEdge and XCLS fail to properly cluster XML documents derived from different
DTDs, although XEdge outperforms XCLS both in the case of homogeneous and
heterogeneous XML documents. However, scalability of the approach has not been
checked.
6.1.3. PSim. Path Similarity (PSim) [Choi et al. 2007] proposes a clustering method
that stores data nodes in an XML document into a native XML storage using path
similarities between data nodes. The proposed method uses path similarities between
data nodes, which reduces the page I/Os required for query processing. According to
our generic framework, PSim presents the following phases.


25:27

—Data Representation. XML documents are modeled as data trees, including both
elements and attributes. No more normalization/transformation process is needed.
—Similarity Computation. PSim uses the Same Path (SP) clusters, where all nodes
(simple objects) with the same absolute path are stored in the same cluster, as the
basic units for similarity computation. The PSim method identifies the absolute path
for each SP cluster and then compares SP clusters utilizing their absolute paths as
identifiers. It computes the path similarity between every SP cluster pair using
the edit distance algorithm. Given two absolute paths P1 = /a1 /a2 .../an and P2 =
/b1 /b2 /.../bm, the edit distance between two paths, δ(P1 , P2 ), can be computed using
a dynamic programming technique. Using the edit distance between two absolute
paths, a path similarity between two SP clusters sp1 and sp2 is computed as follows.
δ(P1 , P2 )
.
max(|P1 ||P2 |)
—Clustering/Grouping. The path similarity matrix is represented as a weighted graph,
named the path similarity graph, where nodes represent the set of SP clusters and
edges connect them with weights. The weight of each edge is the path similarity
between two nodes that are connected by that edge. The greedy algorithm is used to
partition the path similarity graph.
—Evaluation Criteria. To validate the performance of the PSim system, it is evaluated
with 1000 randomly generated path queries over an XML document with 200,000
nodes. The sum of page I/Os required for query processing is used as the performance
criterion. Experimental results show that the PSim clustering method has good performance in query processing, however the proposed method assumes that document
updates are infrequent. Further studies on the clustering method are needed when
updates are frequent.
path similarity(sp1 , sp2 ) = 1 −

6.1.4. XProj. XProj [Aggarwal et al. 2007] introduces a clustering algorithm for XML
documents that uses substructures of the documents in order to gain insight into the
important underlying structures.

—Data Representation. XML documents are modeled as ordered data trees. XProj does
not distinguish between attributes and elements of an XML document, since both are
mapped to the label set. For efficient processing, the preorder depth-first traversal
of the tree structure is used, where each node is represented by the path from the
root node to itself. The content within the nodes is ignored and only the structural
information is used.
—Similarity Computation. XProj makes use of frequent substructures in order to define
similarity among documents. This is analogous to the concept of projected clustering
in multidimensional data, hence the name XProj. In the projected clustering algorithm, instead of using individual XML documents as representatives for partitions,
XProj uses a set of substructures of the documents. The similarity of a document to
a set of structures in a collection is the fraction of nodes in the document that are
covered by any structure in the collection. This similarity can be generalized to similarity between a set of documents and a set of structures by averaging the structural
similarity over the different documents. This gives the base for computing the frequent substructural self-similarity between a set of XML documents. To reduce the
cost of similarity computation, XProj selects the top-K most frequent substructures
of size l.
—Clustering/Grouping. The primary approach is to use a substructural modification of
a partition-based approach, in which the clusters of XML documents are built around
groups of representative substructures. Initially, the sets of XML documents are divided into K partitions with equal size, and the sets of substructure representatives

25:28

A. Algergawy et al.

(a) XML data tree of D1 in Fig. 4(a)

path
Paper.Journal.Author
Paper.Journal.Title
Paper.Journal.page

Answer
John
XML Data Clustering
25-36

(b) tree tuples

node ID
n4
n6
n8

transaction
tr1

set
n4 , n6 , n8

(c) Transaction set

Fig. 12. Transaction representation of an XML document.

are generated by mining frequent substructures from these partitions. These structure representatives are used to cluster the XML documents using the self-similarity
measure.
—Evaluation Criteria. The two external criteria, precision and recall, are used to evaluate the quality of the XProj clustering method. XProj is also compared with the
Chawathe algorithm [Chawathe 1999] and the structural algorithm [Dalamagas
et al. 2006]. The comparison results indicate that both the XProj and the structural
algorithms work well and have a higher precision and recall than Chawathe’s tree
edit distance-based algorithm when dealing with heterogeneous XML documents. In
the case of homogeneous XML documents, XProj outperforms the other two systems.
6.1.5. SemXClust. The Semantic XML Clustering (SemXClust) method [Tagarelli and
Greco 2006] investigates how to cluster semantically related XML documents through
in-depth analysis of content information extracted from textual elements and structural information derived from tag paths. Both kinds of information are enriched with
knowledge provided by a lexical ontology.

—Data Representation. The XML documents are represented as data trees. SemXClust
defines an XML tree tuple as the maximal subtree of the document tree satisfying
the path/answer condition. This means that tree tuples extracted from the same tree
maintain an identical structure while reflecting different ways of associating content
with structure. For efficient processing, especially in the huge amount of available
structured data, a relevant portion is represented by variable-length sequences of
objects with categorical attributes, named transactional data, as shown in Figure 12.
—Similarity Computation. Since SemXClust represents and embeds XML features in
tree tuple items, the notion of similarity between tree tuple items is a weighted sum
function of the similarity between their respective structure and content features.
The structural similarity between two tree tuple items is determined by comparing
their respective tag paths and computing the average similarity between the senses of
the respective best matching tags. Furthermore, the content similarity is determined
using the text extracted from any leaf node of an XML tree. The extracted text is
represented as a bag-of-words model and then subjected to both lexical and semantic
analysis.
—Clustering/Grouping. SemXClust applies a partitional algorithm devised for the
XML transactional domain, called TrK-means [Giannotti et al. 2002]. The clustering
algorithm has two phases: (1) working as a traditional centroid-based method to
compute k + 1 clusters, and (2) recursively splitting the (k + 1)th cluster into a small
number of clusters. SemXClust adapts the TrK-means clustering algorithm focusing
on conceiving suitable notions of proximity among XML transactions. The resulting
algorithm is called XTrK-means.


25:29

(a) XML data tree of D2 and D3

edges
paper → conference
conference → author
conference → title
conference → page
conference → url

entry 1
1
1
1
1
0
DT2

entry 2
1
1
1
0
1
DT3

(b) SG
Fig. 13. S-Graph encoding.

—Evaluation Criteria. The intra- and inter-clustering similarity criteria are used to
validate the quality of the clustering solution. Both criteria are based on the pairwise
similarity between transactions. No comparison with other clustering algorithms has
been done. Furthermore, SemXClust is not tested against clustering scalability.
6.1.6. S-GRACE. S-GRACE [Lian et al. 2004] proposes a hierarchical algorithm for
clustering XML documents based on the structural information in the data. A distance
metric is developed on the notion of the structure graph, which is a minimal summary
of edge containment in the documents. According to our generic framework, S-GRACE
has the following phases.

—Data Representation. An XML document is represented as a directed graph called
the structure graph (or s-graph), where the node set includes the set of all elements
and attributes of the document. The s-graphs of all documents are analyzed and
stored in a data structure called SG. Each entry in the SG contains two information
ﬁelds, as shown in Figure 13: (1) a bit string representing the edges of an s-graph,
and (2) a set containing the IDs of documents whose s-graphs are represented by this
bit string. Therefore, the problem of clustering XML documents is transformed into
clustering a smaller set of bit strings.
—Similarity Computation. S-GRACE computes the distance between all pairs of sgraphs in SG using a distance metric proposed for s-graphs and given by the following
equation.
|sg(DT 1) ∩ sg(DT 2)|
δ(DT 1, DT 2) = 1 −
,
max{|sg(DT 1)|, |sg(Dt2)|}
where |sg(DT 1)| is the number of edges in sg(DT 1), and |sg(DT 1) ∩ sg(DT 2)| is
the set of common edges of sg(DT 1) and sg(DT 2). Instead of using a pure distancebased algorithm, S-GRACE, as in ROCK [Guha et al. 2000], considers the number
of common neighbors. The distance between all pairs of s-graphs in SG is computed
and stored in a distance array.
—Clustering/Grouping. S-GRACE is a hierarchical clustering algorithm, which applies
ROCK [Guha et al. 2000] on the s-graphs. Initially, each entry in SG forms a separate
cluster. The clustering technique builds two heaps—a local heap q[i] for each cluster
i and a global heap Q containing all clusters. The clustering algorithm iterates until
α × K clusters remain in the global heap, where α is a small number controlling the
merging process.
—Evaluation Criteria. The quality of S-GRACE is evaluated using nonstandard criteria, such as the closeness between cluster and the average similarity over all
pairs of clusters, while the scalability is validated using the response time. The time

25:30

A. Algergawy et al.

complexity of the S-GRACE method is O(m2 log m), where m is the number of distinct
s-graphs in SG. Experimental results using both synthetic and real-world data sets
show that S-GRACE is effective in identifying clusters, however it does not scale
well. The system needs 4500 seconds to cluster a set of XML documents with a size
of 200KB.
6.1.7. SumXClust. The Clustering XML documents exploiting structural summaries
(SumXClust)7 method [Dalamagas et al. 2006] represents a clustering algorithm for
grouping structurally similar XML documents using the tree structural summaries to
improve the performance of the edit distance calculation. To extract structural summaries, SumXClust performs both nesting reduction and repetition reduction.

—Data Representation. XML documents are represented as ordered data trees. SumXClust proposes the use of compact trees, called tree structural summaries that have
minimal processing requirements instead of the original trees representing XML
documents. It performs nesting reduction and repetition reduction to extract structural summaries for data trees. For nesting reduction, the system traverses the data
tree using preorder traversal to detect nodes that have an ancestor with the same
label in order to move up their subtrees. For repetition reduction, it traverses the
data tree using preorder traversal ignoring already existing paths and keeping new
ones, using a hash table.
—Similarity Computation. SumXClust proposes a structural distance S between two
XML documents represented as structural summaries. The structural distance is
deﬁned as:
δ(DT1 , DT2 )
S(DT1 , DT2 ) =
,
δ (DT1 , DT2 )
where δ (DT1 , DT2 ) is the cost to delete all nodes from DT1 and insert all nodes from
DT2 . To determine edit distances between structural summaries δ(DT1 , DT2 ), the
system uses a dynamic programming algorithm that is close to Chawathe [1999].
—Clustering/Grouping. SumXClust implements a hierarchical clustering algorithm
using Prim’s algorithm for computing the minimum spanning tree. It forms a fully
connected graph with n nodes from n structural summaries. The weight of an edge
corresponds to the structural distance between the nodes that this edge connects.
—Evaluation Criteria. The performance, as well as the quality, of the clustering results
is tested using synthetic and real data. To evaluate the quality of the clustering
results, two external criteria, precision and recall, have been used, while the response
time is used to evaluate the efﬁciency of the clustering algorithm. SumXClust is
also compared with the Chawathe algorithm [Chawathe 1999]. Experimental results
indicate that with or without summaries, the SumXClust algorithm shows excellent
clustering quality, and improved performance compared to Chawathe’s.
6.1.8. VectXClust. Clustering XML documents based on the vector representation for
documents (VectXClust)8 [Yang et al. 2005b] proposes the transformation of treestructured data into an approximate numerical multidimensional vector, which encodes
the original structure information.

—Data Representation. XML documents are modeled as ordered data trees, which are
then transformed into full binary trees. A full binary tree is a binary tree in which
each node has exactly one or two children. The binary trees are then mapped to
numerical multidimensional vectors, called binary branch vectors, where the features
7 We
8 We

give the tool this name for easier reference.



25:31

of the vectors retain the structural information of the original trees. The binary
branch vector can be extended to the characteristic vector, which includes all the
elements in the q-level binary branch space.
—Similarity Computation. Based on the vector representation, VectXClust defines a
new distance, called the binary branch distance, of the tree structure as the L1
distance between the vector images of two trees. The methodology is to embed the
defined distance function, that is, the lower bound of the actual tree edit distance into
a filter-and-refine framework. This allows filtering out very dissimilar documents and
computing the tree edit distance only with a restricted number of documents.
—Clustering/Grouping. In fact, VectXClust does not conduct any clustering step. However, the measured distances between XML documents can be used later for any
clustering process.
6.1.9. DFTXClust. Clustering XML documents based on the time series representation
for XML documents (DFTXClust)9 [Flesca et al. 2005] proposes the linearization of
the structure of each XML document, by representing it as a numerical sequence, and
then comparing such sequences through the analysis of their frequencies. The theory
of the Discrete Fourier Transform is exploited to compare the encoded documents in
the domain of frequencies.

—Data Representation. DFTXClust is only interested in the structure of XML documents, hence it limits its attention to start tags and end tags. Each tag instance is
denoted by a pair composed by its unique identifier and its textual representation.
Moreover, it is necessary to consider the order of appearance of tags within a document. To this end, given an XML document docx , the authors define its skeleton
as the sequence of all tag instances appearing with docx . The system relies on the
effective encoding of the skeleton of an XML document into a time series summarizing its features. For this purpose, the authors develop two encoding functions. A
tag encoding function, which assigns a real value to each tag instance, and a document encoding function, which associates a sequence of reals with the skeleton of
the document. In order to measure the impact of encoding functions in detecting
dissimilarities among documents, several encoding schemes have been proposed.
—Similarity Computation. Based on the preorder visit of the XML document, starting
at initial time t0 , the authors assume that each tag instance occurs after a fixed time
interval . The total time spent to visit the document is n , where n is the size of
tags(docx ). During the visit, the system produces an impulse that depends on a particular encoding tag function and the document encoding function. Instead of using
Time Warping to compare sequences, Discrete Fourier Transform (DFT) is used to
transform document signals into vectors whose components correspond to frequencies in the interval [−0.5, 0.5]. Based on this representation, the authors define a
metric distance called the Discrete Fourier Transform distance as the approximation of the difference of the magnitudes of the DFTs of the two encoded documents.
The result of this phase is a similarity matrix representing the degree of structural
similarity for each pair of XML documents in the data set.
6.2. Clustering XML Schema Prototypes

6.2.1. XClust. XClust [Lee et al. 2002] proposes an approach of clustering the DTDs of
XML data sources to effectively integrate them. It is defined by using the phases of the
generic framework as follows.

9 We



25:32

A. Algergawy et al.

—Data Representation. DTDs are analyzed and represented as unordered data trees.
To simplify schema matching, a series of transformation rules are used to transform
auxiliary OR nodes (choices) to AND nodes (sequences), which can be merged. XClust
makes use of simple objects, their features, and relationships between them.
—Similarity Computation. To compute the similarity of two DTDs, XClust is based on
the computation of the similarity between simple objects in the data trees. To this end,
the system proposes a method that relies on the computation of semantic similarity
exploiting semantic features of objects, and on the computation of the structure and
context similarity exploiting relationships between objects. The output of this phase
is the DTD similarity matrix.
—Clustering/Grouping. The DTD similarity matrix is exploited by a hierarchical clustering algorithm to group DTDs into clusters. The hierarchical clustering technique
can guide and enhance the integration process, since the clustering technique starts
with clusters of single DTDs and gradually adds highly similar DTDs to these
clusters.
—Evaluation Criteria. Since the main objective of XClust is to develop an effective
integration framework, it uses criteria to quantify the goodness of the integrated
schema. No study concerning the clustering scalability has been done.
6.2.2. XMine. XMine [Nayak and Iryadi 2007] introduces a clustering algorithm based
on measuring the similarity between XML schemas by considering the semantics, as
well as the hierarchical structural similarity of elements.

—Data Representation. Each schema is represented as an ordered data tree. A simplification analysis of the data (schema) trees is then performed in order to deal with
the nesting and repetition problems using a set of transformation rules similar to
those in Lee et al. [2002]. XMine handles both the DTD and XSD schemas, and,
like XClust, makes use of simple objects, their features, and relationships, between
objects.
—Similarity Computation. XMine determines the schema similarity matrix through
three components. (1) The element analyzer, determines linguistic similarity by comparing each pair of elements of two schemas primarily based on their names. It
considers both the semantic relationship as found in the WordNet thesaurus and
the syntactic relationship using the string edit distance function. (2) The maximally
similar paths finder identifies paths and elements that are common and similar
between each pair of tree schemas based on the assumption that similar schemas
have more common paths. Moreover, it adapts the sequential pattern mining algorithm [Srikant and Agrawal 1996] to infer the similarity between elements and
paths. (3) The schema similarity matrix processor, in which the similarity matrix between schemas is computed based on the preceding measured criteria. This matrix
becomes the input to the next phase.
—Clustering/Grouping. The constrained hierarchical agglomerative clustering algorithm is used to group similar schemas exploiting the schema similarity matrix.
XMine makes use of the wCluto10 Web-enabled data clustering applications to form
a hierarchy of schema classes.
—Evaluation Criteria. XMine is tested using real-world schemas collected from different domains. To validate the quality of XMine, the standard criteria including
FScore, intra- and inter-clustering similarity have been used. The evaluation shows
that XMine is effective in clustering a set of heterogenous XML schemas. However,
XMine is not subject to any scalability test and a comparison with another clustering
method is missing.
10 http://cluto.ccgb.umn.edu/cgi-bin/wCluto.



25:33

6.2.3. PCXSS. The Progressively Clustering XML by Semantic and Structural Similarity (PCXSS) [Nayak and Tran 2007] method introduces a fast clustering algorithm
based on a global criterion function CPSim (Common Path Coefficient) that progressively measures the similarity between an XSD and existing clusters, ignoring the need
to compute the similarity between two individual XSDs. The CPSim is calculated by
considering the structural and semantic similarity between objects.

—Data Representation. Each schema is represented as an ordered data tree. A simplification analysis of the data (schema) trees is then performed in order to deal with the
composite elements. The XML tree is then decomposed into path information called
node paths. A node path is an ordered set of nodes from the root node to a leaf node.
—Similarity Computation. PCXSS uses the CPSim to measure the degree of similarity
of nodes (simple objects) between node paths. Each node in a node path of a data tree
is matched with the node in a node path of another data tree, and then aggregated
to form the node path similarity. The node similarity between nodes of two paths is
determined by measuring the similarity of its features such as name, data type, and
cardinality constraints.
—Clustering/Grouping. PCXSS is motivated by the incremental clustering algorithms.
It first starts with no clusters. When a new data tree comes in, it is assigned to a new
cluster. When the next data tree comes in, it matches it with the existing cluster.
PCXSS rather progressively measures the similarities between a new XML data tree
and existing clusters by using the common path coefficient (CPSim) to cluster XML
schemas incrementally.
—Evaluation Criteria. PCXSS is evaluated using the standard criteria. Both the quality and the efficiency of PCXSS are validated. Furthermore, the PCXSS method is
compared with a pairwise clustering algorithm, namely wCluto [Zhao and Karypis
2002b]. The results show that the increment in time with the increase of the size of
the data set is less with PCXSS in comparison to the wCluto pairwise method.
6.2.4. SeqXClust. The Sequence matching approach to cluster XML schema (SeqXClust) [Algergawy et al. 2008b] method introduces an efficient clustering algorithm
¨
based on the sequence representation of XML schema using the Prufer encoding
method.

—Data Representation. Each schema is represented as an ordered data tree. A simplification analysis of the data (schema) trees is then performed in order to deal with
the composite elements. Each data tree is then represented as a sequence, using a
¨
modified Prufer encoding method. The semantic information of the data tree is cap¨
tured by Label Prufer sequences (LPSs), while the structure information is captured
¨
by Number Prufer sequences (NPSs).
—Similarity Computation. SeqXClust uses a sequence-based matching approach to
measure the similarity between data trees. It first measures the similarity between
nodes (simple objects) exploiting their features such as name, data type, and relationships between them, such as the node context. The output of this phase is the
schema similarity matrix.
—Clustering/Grouping. SeqXClust, like XMine, makes use of the constrained hierarchical agglomerative clustering algorithm to group similar schemas exploiting the
schema similarity matrix.
—Evaluation Criteria. The SeqXClust method is tested using real-world data sets
collected from different domains. The quality of the method is evaluated using the
standard evaluation criteria. However, neither a scalability study nor a comparison
evaluation has been conducted.


25:34

A. Algergawy et al.

6.3. Comparison of Discussed Clustering Techniques

Table II shows how the discussed XML data clustering prototypes fit the classification criteria introduced in this article. Furthermore, the table shows which part of
the solution is covered by which prototypes, thereby supporting a comparison of the
illustrated approaches. It also indicates the supported clustered data types, the internal data representations, the proximity computation methodologies, the exploited
clustering algorithms, and the performance evaluation aspects.
The table shows that all systems support data representation, most of them in the
form of data trees, only one, S-GRACE, as a directed graph, and two, VectXClust and
DFTXClust, as a vector representation. XML schema-based prototypes perform an
extra normalization step on the schema trees, while some document-based prototypes
extend document trees to handle them efficiently. Moreover, schema-based prototypes
make use of schema matching-based approaches to compute the schema similarity
matrix. Tree matching-based approaches are used for document-based prototypes. On
the other hand, metric distances such as L1 (Manhattan), cosine measures, and Discrete Fourier Transform distance, are used for vector-based approaches. To represent
the similarity across schemas/documents, some prototypes use the array (matrix) data
structure and one makes use of fully connected graphs such as SumXClust.
It is worth noting that some prototypes first complete the similarity computation
process and then apply the clustering algorithm, while others perform similarity
computation during clustering. For example, XClust performs interschema similarity
computation to get the schema similarity matrix, and then applies the hierarchical
clustering algorithm to form a hierarchy of schema classes, while XCLS computes the
level similarity to quantify the structural similarity between an XML document and
existing clusters, and groups the XML document to the cluster with the maximum level
similarity.
From the performance point of view, nearly all XML data clustering prototypes
evaluate their performance according to the quality aspect considering the internal
measures, such as SemXClust, or the external measures, such as SumXClust and
XProj, or both, such as XMine and XCLS. The XClust prototype evaluates its quality
according to the quality of the integrated schema based on the context of the application domain. To evaluate their scalability, the prototypes use the time response
as a measure of efficiency. Since most of these prototypes make use of pairwise
clustering algorithms, their time complexity is at least O(N 2 ), where N is the number
of elements of XML data. This is infeasible for large amounts of data. However, no
evaluation work has been done considering a trade-off between the two performance
aspects.
Finally, clustering XML data is useful in numerous applications. Table II shows that
some prototypes have been developed and implemented for specific XML applications,
such as XClust for DTD integration and S-GRACE for XML query processing, while
other prototypes have been developed for generic applications, such as XMine and
XCLS for XML management in general.
7. SUMMARY AND FUTURE DIRECTIONS

Clustering XML data is a basic problem in many data application domains such
as Bioinformatics, XML information retrieval, XML data integration, Web mining,
and XML query processing. In this article, we conducted a survey about XML
data clustering methodologies and implementations. In particular, we proposed a
classification scheme for these methodologies based on three major criteria: the type
of clustered data (clustering either XML documents or XML schemas), the similarity
measure approach (either the tree similarity-based approach or the vector-based

Tree matching


Efﬁciency
evaluation
Application
area

Quality criteria

Cluster
representation

Vector
based
Similarity
representation
Clustering
algorithm

Similarity
approach

XClust

XMine

PCXSS

—

—

—
general XML
management

—

hierarchy of
schema classes
(dendrogram)
quality of
integrated
schema

hierarchy of
schema classes
(dendrogram)
inter &
intra-cluster
FScore

DTD similarity
matrix
hierarchical

schema similarity
matrix
hierarchical
agglomerative

—

—

XML data
integration

response
time
general XML
management

inter &
intra-cluster
FScore

K-clusters

incremental

—

—

—

DTD
DTD&
XSD
schema
XSD schema
schema
rooted unordered
rooted ordered
rooted ordered
labeled tree
labeled tree
labeled tree
elements,
elements,
attributes,
attributes,
node paths
paths
paths
name,
linguistic,
name,
Schema matching
cardinality,
hierarchical structure, data type,
path context
path similarity
constraint

Clustered
data type
Data
representation
Exploited
objects

XXXPrototypes
XX
Criteria
X

—

level
similarity

—

XML
documents
rooted ordered
labeled tree
level
information
elements

XCLS

—

edge
similarity
—

path
similarity

—

paths

edge information

—

XML
documents
rooted
labeled tree

PSim

XML
documents
rooted ordered
labeled tree

XEdge

—

XML
documents
rooted
labeled tree
subtrees
(tree tuple)

SemXClust

—

—

frequent
content &
substructure structural similarity

—

substructures

XML
documents
rooted ordered
labeled tree

XProj

schema similarity
edge similarity
path similarity
—
—
—
matrix
matrix
matrix
constrained
graph partitioning partition-base
partitional
incremental
K-means
hierarchical
greedy algorithm
approach
XTrK-means
agglomerative
algorithm
hierarchy of
schema classes
K-clusters
K-clusters
K-clusters
K-clusters
K-clusters
(dendrogram)
inter &
external &
Precision
intra-cluster internal measures
FScore
—
recall
inter & intra-cluster
FScore
purity, entropy,
FScore inter &
intra cluster
time & space
page I/O
response
—
—
—
complexity
time
general XML
general XML
general XML
XML storage
XML data
XML data
management
management
management
for query
mining
mining
processing

—

—

XSD
schema
rooted ordered
labeled tree
elements,
attributes,
node context
name,
data type,
node context

SeqXClust

Table II. Features of XML Data Clustering Approaches
SumXClust

VectXClust

response
time
XML query
processing

closeness
standard
deviation
outlier ratio

K-clusters

distance
array
hierarchical
ROCK
approach

—

distance
metric

—

response
time
hierarchical
structural
management

precision
recall

K-clusters

fully connected
graph
single link
hierarchical

—

structural
distance

—

response
time
Query
processing

—

—

—

—

binary branch
distance

—

—

XML
XML
XML
documents
documents
documents
directed
rooted ordered
binary
graph
labeled tree branch vector
whole XML tree whole XML tree whole XML
(s-graph)
(structural
tree
summaries)

S-GRACE

Storage &
retrieval of
large documents

—

—

—

—

—

DFT
distance

—

—

XML
documents
time
series
whole XML
tree

DFTXClust

25:35

25:36

A. Algergawy et al.

approach), and the clustering algorithm. In order to base the comparison between
their implementations, we also proposed a generic clustering framework.
In this article, we devised the main current approaches for clustering XML data relying on similarity measures. The discussion also pointed out new research trends that
are currently under investigation. In the remainder, we sum up interesting research
issues that still deserve attention from the research community classified according to
four main directions: semantics, performance, trade-off between clustering quality and
clustering scalability, and application context directions.
—Semantics. More semantic information should be integrated in clustering techniques
in order to improve the quality of the generated clusters. Examples of semantic information are: schema information that can be associated with documents or extracted
from a set of documents; ontologies and thesauri, that can be exploited to identify
similar concepts; domain knowledge, that can be used within a specific context to
identify as similar concepts that in general are retained differently (for example,
usually papers in proceedings are considered different from journal papers, however
a paper in the VLDB proceedings has the same relevance as a paper in an important
journal). The semantics of the data to be clustered can also be exploited for the selection of the measure to be employed. Indeed, clustering approaches mainly rely on
well known fundamental measures (e.g., tree edit distance, cosine, and Manhattan
distances) that are general purpose and cannot be easily adapted to the characteristics of data. A top-down or bottom-up strategy can be followed for the specification
of these kinds of functions. Following the top-down strategy, a user interface can be
developed offering the user a set of building block similarity measures (including the
fundamental ones) that can be applied on different parts of XML data. Users can
compose these measures in order to obtain a new one that is specific for their needs.
The user interface should offer the possibility of comparing the effectiveness of different measures and choosing the one that best suits their needs. In this strategy, the
user is in charge of specifying the measure to be used for clustering. In the context of
approximate retrieval of XML documents, the multisimilarity system arHex [Sanz
et al. 2006] has been proposed. Users can specify and compose their own similarity
measures and compare the retrieved results depending on the applied measures.
Analogous approaches should be developed for clustering XML data, also for incrementally checking the correctness of the specified measure, and for changing the
clustering algorithm.
Following the bottom-up strategy, the characteristics of the data that need to be
clustered are considered to point out the measure (or the features of the measure)
that should return the best result (Muller et al. [2005b, 2005a] developed a similar
approach in other contexts). Exploiting functions that extract different kinds of information from the data (e.g., element tags, relevant words of elements, attributes,
links), the features of the similarity measures to be applied on the data can be
pointed out. For example, if attributes seldom occur in a collection of documents, a
similarity measure that relies on attributes is not really useful; by contrast, if the
internal structures of the documents are mostly identical, a content-based measure
is expected. In this strategy, therefore, the data characteristics suggest the measure
to apply for clustering. Sanz et al. [2008] propose the use of entropy-based measures
to evaluate the level of heterogeneity of different aspects of documents. This information can be used to choose the more adequate indexing structure, and to find out
whether it is better to use an exact or approximate retrieval approach. The integration of the two strategies and their employment in XML data clustering is thus a
very interesting research direction.


25:37

Finally, most approaches are based on equality comparisons concerning the evaluation of the similarity at single elements/nodes. A more semantic approach, relying
on ontologies and thesauri for allowing multilingual data handling and concept-based
clustering, would certainly be useful.
—Performance. Another research issue to be faced, is the performance of the similarity evaluation process. When the size and the number of the data to be clustered
increase, the current approaches do not scale well. A possible solution that needs
to be investigated is to employ two measures in the evaluation of similarity (in
the spirit of Yang et al. [2005a]). The first should be able to quickly identify the
documents that are dissimilar in order to identify data that presumably belong
to the same cluster. Then, using a finer (and time-consuming) function, the previous results should be revised on a reduced number of documents/schemas. The
key issue is the choice of the two measures that should be compatible, that is,
the two measures should return comparable results even if with different performance (currently in Yang et al. [2005a] a factor 5 between the two measures that
is too high). Orthogonally to this approach, a ranking on the performance of the
approaches proposed so far should eventually be provided bound to their theoretical
complexity.
—Trade-off Between Clustering Quality and Clustering Efficiency. Many real-world
problems such as the XML data clustering problem, involve multiple measures of
performance, which should be optimized simultaneously. Optimal performance according to one objective, if such an optimum exists, often implies unacceptably low
performance in one or more of the other objective dimensions, creating the need for
a compromise to be reached. In the XML data clustering problem, the performance
of a clustering system involves multiple aspects, among them clustering quality and
clustering efficiency. Optimizing one aspect, for example clustering quality, will affect the other aspects such as efficiency, and so on. Hence, we need a compromise
between them, and we could consider the trade-off between quality and efficiency
of the clustering result as a multiobjective optimization problem [Algergawy et al.
2008a]. In practice, multiobjective problems have to be reformulated as a single objective problem. This representation enables us to obtain a combined measure as a
compromise between them.
A critical scenario in which a trade-off between clustering quality and clustering
efficiency should be taken into account can be stated as follows. Assume we have a set
of XML data to be clustered, and we have two clustering systems, A and B. System
A is more effective than system B, while system B is more efficient than system A.
The question arising here is which system will be used to solve the given problem?
—Application Context. Another interesting research direction is the employment of
the data context in the evaluation of similarity. The context can be any information about the data that is not contained in the file. For example, the authors
of the data, the creation date, the geospatial position of the file. This information can be integrated in the similarity measure in order to obtain more precise
results.
ACKNOWLEDGMENTS
We wish to thank Stefanie Quade for improving the quality of the article. We are grateful to the anonymous
reviewers for significantly improving the article by their detailed and deeply knowledgeable comments.

REFERENCES
AGGARWAL, C. C., TA, N., WANG, J., FENG, J., AND ZAKI, M. J. 2007. XProj: a framework for projected structural
clustering of XML documents. In Proceedings of the 13th ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining (KDD). 46–55.


25:38

A. Algergawy et al.

ALGERGAWY, A., SCHALLEHN, E., AND SAAKE, G. 2008a. Combining effectiveness and efficiency for schema matching evaluation. In Proceedings of the 1st International Workshop Model-Based Software and Data Integration (MBSDI’08). 19–30.
ALGERGAWY, A., SCHALLEHN, E., AND SAAKE, G. 2008b. A schema matching-based approach to XML schema
clustering. In Proceedings of the 10th International Conference on Information Integration and WebBased Applications & Services (iiWAS). 131–136.
ALGERGAWY, A., SCHALLEHN, E., AND SAAKE, G. 2009. Improving XML schema matching using Prufer sequences.
Data Knowl. Eng. 68, 8, 728–747.
ANDREOPOULOS, B., AN, A., WANG, X., AND SCHROEDER, M. 2009. A roadmap of clustering algorithms: finding a
match for a biomedical application. Briefings Bioinform. 10, 3, 297–314.
ANTONELLIS, P., MAKRIS, C., AND TSIRAKIS, N. 2008. XEdge: Clustering homogeneous and heterogeneous XML
documents using edge summaries. In Proceedings of the ACM Symposium on Applied Computing (SAC).
1081–1088.
BAEZA-YATES, R. AND RIBEIRO-NETO, B. 1999. Modern Information Retrieval. ACM Press/Addison-Wesley.
BATINI, C., LENZERINI, M., AND NAVATHE, S. 1986. A comparative analysis of methodologies for database schema
integration. ACM Comput. Surv. 18, 4, 323–364.
BERKHIN, P. 2002. Survey of clustering data mining techniques. 10.1.1.145.895.pdf.
BERTINO, E. AND FERRARI, E. 2001. XML and data integration. IEEE Internet Comput. 5, 6, 75–76.
BERTINO, E., GUERRINI, G., AND MESITI, M. 2004. A matching algorithm for measuring the structural similarity
between an XML document and a DTD and its applications. Inform. Syst. 29, 1, 23–46.
BERTINO, E., GUERRINI, G., AND MESITI, M. 2008. Measuring the structural similarity among XML documents
and DTDs. Intell. Inform. Syst. 30, 1, 55–92.
BILLE, P. 2005. A survey on tree edit distance and related problems. Theoret. Comput. Sci. 337, 1-3,
217–239.
BOLSHAKOVA, N. AND CUNNINGHAM, P. 2005. cluML: a markup language for clustering and cluster validity
assessment of microarray data. Tech. rep. TCD-CS-2005-23, The University of Dublin.
BONIFATI, A., MECCA, G., PAPPALARDO, A., RAUNICH, S., AND SUMMA, G. 2008. Schema mapping verification:
the spicy way. In Proceedings of the 11th International Conference on Extending Database Technology
(EDBT). 85–96.
BOUKOTTAYA, A. AND VANOIRBEEK, C. 2005. Schema matching for transforming structured documents. In Proceedings of the ACM Symposium on Document Engineering (DocEng). 101–110.
BOURRET, R. 2009. XML database products. http://www.rpbourret.com/xml/XMLDatabaseProds.htm.
BUTTLER, D. 2004. A short survey of document structure similarity algorithms. In Proceedings of the International Conference on Internet Computing. 3–9.
CAN, F. 1993. Incremental clustering for dynamic information processing. ACM Trans. Inform. Syst. 11, 2,
143–164.
CANDILLIER, L., TELLIER, I., AND TORRE, F. 2005. Transforming XML trees for efficient classification and clustering. In Proceedings of the INEX Workshop on Element Retrieval Methodology. 469–480.
CERAMI, E. 2005. XML for Bioinformatics. Springer New York.
CHARIKAR, M., CHEKURI, C., FEDER, T., AND MOTWANI, R. 2004. Incremental clustering and dynamic information
retrieval. SIAM J. Comput. 33, 6, 1417–1440.
CHAWATHE, S. S. 1999. Comparing hierarchical data in external memory. In Proceedings of the 25th International Conference on Very Large Data Bases (VLDB). 90–101.
CHOI, I., MOON, B., AND KIM, H.-J. 2007. A clustering method based on path similarities of XML data. Data
Knowl. Eng. 60, 361–376.
COHEN, W. W., RAVIKUMAR, P., AND FIENBERG, S. E. 2003. A comparison of string distance metrics for namematching tasks. In Proceedings of the IJCAI-03 Workshop on Information Integration on the Web (IIWeb).
73–78.
CRISTIANINI, N., SHAWE-TAYLOR, J., AND LODHI, H. 2002. Latent semantic kernels. J. Intell. Inform. Syst. 18, 2-3,
127–152.
DALAMAGAS, T., CHENG, T., WINKEL, K.-J., AND SELLIS, T. 2006. A methodology for clustering XML documents by
structure. Inform. Syst. 31, 187–228.
DO, H. H. AND RAHM, E. 2002. COMA- A system for flexible combination of schema matching approaches. In
Proceedings of the 28th International Conference on Very Large Data Bases (VLDB). 610–621.
DOAN, A., MADHAVAN, J., DOMINGOS, P., AND HALEVY, A. 2004. Handbook on Ontologies. Springer, 385–404.



25:39

DORNELES, C. F., HEUSER, C. A., LIMA, A. E. N., DA SILVA, A. S., AND DE MOURA, E. S. 2004. Measuring similarity
between collection of values. In Proceedings of the 6th ACM CIKM International Workshop on Web
Information and Data Management (WIDM). 56–63.
EISEN, M. B., SPELLMAN, P. T., BROWN, P. O., AND BOTSTEIN, D. 1998. Cluster analysis and display of genome-wide
expression patterns. Proc. Natl. Acad. Sci. USA 95, 14863–14868.
FLESCA, S., MANCO, G., MASCIARI, E., PONTIERI, L., AND PUGLIESE, A. 2005. Fast detection of XML structural
similarity. IEEE Trans. Knowl. Data Eng. 17, 2, 160–175.
FLORESCU, D. AND KOSSMANN, D. 1999. Storing and querying XML data using an RDMBS. IEEE Data Eng.
Bull. 22, 3, 27–34.
GIANNOTTI, F., GOZZI, C., AND MANCO, G. 2002. Clustering transactional data. In Proceedings of the European
Conference on Principles of Data Mining and Knowledge Discovery (PKDD). 175–187.
GIUNCHIGLIA, F., YATSKEVICH, M., AND SHVAIKO, P. 2007. Semantic matching: Algorithms and implementation.
J. Data Semantics 9, 1–38.
GOU, G. AND CHIRKOVA, R. 2007. Efficiently querying large XML data repositories: A survey. IEEE Trans.
Knowl. Data Eng. 19, 10, 1381–1403.
GUERRINI, G., MESITI, M., AND SANZ, I. 2007. An Overview of Similarity Measures for Clustering XML Documents. Web Data Management Practices: Emerging Techniques and Technologies. IDEA GROUP.
GUHA, S., RASTOGI, R., AND SHIM, K. 1998. CURE: An efficient clustering algorithm for large databases. In
Proceedings of the ACM SIGMOD International Conference on Management of Data (SIGMOD). 73–84.
GUHA, S., RASTOGI, R., AND SHIM, K. 2000. ROCK: A robust clustering algorithm for categorical attributes.
Inform. Syst. 25, 5, 345–366.
HANISCH, D., ZIMMER, R., AND LENGAUER, T. 2002. ProML-the protein markup language for specification of
protein sequences, structures and families. Silico Biol. 2, 3, 313–324.
HUANG, Z. 1998. Extensions to the k-means algorithm for clustering large data sets with categorical values.
Data Min. Knowl. Discov. 2, 3, 283–304.
HUCKA, M., FINNEY, A., SAURO, H. M., ET AL. 2003. The systems biology markup language (SBML): A medium
for representation and exchange of biochemical network models. Bioinformatics 19, 4, 524–531.
JAIN, A., MURTY, M., AND FLYNN, P. 1999. Data clustering: A review. ACM Comput. Surv. 31, 3, 264–323.
JEONG, B., LEE, D., CHO, H., AND LEE, J. 2008. A novel method for measuring semantic similarity for XML
schema matching. Expert Syst. Appl. 34, 1651–1658.
JEONG, J., SHIN, D., CHO, C., , AND SHIN, D. 2006. Structural similarity mining in semi-structured microarray
data for efficient storage construction. In Proceedings of the Workshop on the Move to Meaningful Internet
Systems, Lecture Notes in Computer Science, vol. 4277.
KADE, A. M. AND HEUSER, C. A. 2008. Matching XML documents in highly dynamic applications. In Proceedings
of the ACM Symposium on Document Engineering (DocEng). 191–198.
LE, D. X., RAHAYU, J. W., AND PARDEDE, E. 2006. Dynamic approach for integrating web data warehouses. In
Proceedings of the International Conference on Computational Science and Its Applications (ICCSA).
207–216.
LEE, D. AND CHU, W. W. 2000. Comparative analysis of six XML schema languages. SIGMOD Record 9, 3,
76–87.
LEE, M. L., YANG, L. H., HSU, W., AND YANG, X. 2002. Xclust: Clustering XML schemas for effective integration.
In Proceedings of the ACM CIKM International Conference on Information and Knowledge Management
(CIKM). 292–299.
LEE, Y., SAYYADIAN, M., DOAN, A., AND ROSENTHAL, A. 2007. Etuner: tuning schema matching software using
synthetic scenarios. VLDB J. 16, 1, 97–132.
LEITO, L., CALADO, P., AND WEIS, M. 2007. Structure-based inference of XML similarity for fuzzy duplicate
detection. In Proceedings of the 16th ACM Conference on Information and Knowledge Management,
(CIKM). 293–302.
LEUNG, H.-P., CHUNG, F.-L., AND CHAN, S. C.-F. 2005. On the use of hierarchical information in sequential
mining-based XML document similarity computation. Knowl. Inform. Syst. 7, 4, 476–498.
LI, W. AND CLIFTON, C. 2000. SemInt: A tool for identifying attribute correspondences in heterogeneous
databases using neural networks. Data Konwl. Eng. 33, 49–84.
LIAN, W., LOK CHEUNG, D. W., MAMOULIS, N., AND YIU, S.-M. 2004. An efficient and scalable algorithm for
clustering XML documents by structure. IEEE Trans. Knowl. Data Eng. 16, 1, 82–96.
LIANG, W. AND YOKOTA, H. 2005. LAX: An efficient approximate XML join based on clustered leaf nodes
for XML data integration. In Proceedings of the 22nd British National Conference on Databases.
82–97.


25:40

A. Algergawy et al.

LIANG, W. AND YOKOTA, H. 2006. SLAX: An improved leaf-clustering-based approximate XML join algorithm
for integrating XML data at subtree classes. IPSJ Digital Courier 2, 382–392.
MADHAVAN, J., BERNSTEIN, P. A., AND RAHM, E. 2001. Generic schema matching with cupid. In Proceedings of
27th International Conference on Very Large Data Bases (VLDB’01). 49–58.
¨
MANNING, C. D., RAGHAVAN, P., AND SCHUTZE, H. 2008. Introduction to Information Retrieval. Cambridge University Press.
MELNIK, S., GARCIA-MOLINA, H., AND RAHM, E. 2002. Similarity ﬂooding: A versatile graph matching algorithm
and its application to schema matching. In Proceedings of the 18th International Conference on Data
Engineering (ICDE).
MELTON, J. AND BUXTON, S. 2006. Querying XML: XQuery, XPath, and SQL/XML in Context. Morgan
Kaufmann/Elsevier.
MULLER, T., SELINSKI, S., AND ICKSTADT, K. 2005a. Cluster analysis: A comparison of different similarity measures for SNP data. In Proceedings of the Second Joint Meeting of the Institute of Mathematical Statistics
and International Society for Bayesian Analysis (IMS-ISBA).
MULLER, T., SELINSKI, S., AND ICKSTADT, K. 2005b. How similar is it? towards personalized similarity measures
in ontologies. In 7. International Tagung Wirschaftinformatik.
NAYAK, R. 2008. Fast and effective clustering of XML data using structural information. Knowl. Inform.
Syst. 14, 2, 197–215.
NAYAK, R. AND IRYADI, W. 2007. XML schema clustering with semantic and hierarchical similarity measures.
Knowl.-Based Syst. 20, 336–349.
NAYAK, R. AND TRAN, T. 2007. A progressive clustering algorithm to group the XML data by structural and
semantic similarity. Inter. J. Patt. Recog. Artif. Intell. 21, 4, 723–743.
NIERMAN, A. AND JAGADISH, H. V. 2002. Evaluating structural similarity in XML documents. In Proceedings of
the 5th International Workshop on the Web and Databases (WebDB). 61–66.
ORDONEZ, C. AND OMIECINSKI, E. 2004. Efﬁcient disk-based k-means clustering for relational databases. IEEE
Trans. Knowl. Data Eng. 16, 8, 909–921.
PAL, S., TALWAR, V., AND MITRA, P. 2002. Web mining in soft computing framework: Relevance, state of the art
and future directions. IEEE Trans. Neural Netw. 13, 5, 1163–1177.
RAFIEI, D., MOISE, D. L., AND SUN, D. 2006. Finding syntactic similarities between XML documents. In Proceedings of the 17th International Conference on Database and Expert Systems Applications (DEXA).
512–516.
RAHM, E. AND BERNSTEIN, P. A. 2001. A survey of approaches to automatic schema matching. VLDB J. 10, 4,
334–350.
SALEEM, K., BELLAHSENE, Z., AND HUNT, E. 2008. PORSCHE: Performance oriented schema mediation. Inform.
Syst. 33, 7-8, 637–657.
SALTON, G., WONG, A., AND YANG, C. 1975. A vector space model for automatic indexing. Comm. ACM 18, 11,
613–620.
SANZ, I., MESITI, M., BERLANGA, R., AND GUERRINI, G. 2008. An entropy-based characterization of heterogeneity
of XML collections. In Proceedings of the 3rd International Workshop on XML Data Management Tools
and Techniques, (XANTEC). 238–242.
SANZ, I., MESITI, M., GUERRINI, G., AND BERLANGA, R. 2006. Arhex: An approximate retrieval system for highly
heterogeneous XML document collections. In Proceedings of the 10th International Conference on Extending Database Technology (EDBT). 1186–1189.
SASSON, O., LINIAL, N., AND LINIAL, M. 2002. The metric space of proteins-comparative study of clustering
algorithms. Bioinformatics 18, 1, 14–21.
SHANMUGASUNDARAM, J., TUFTE, K., HE, G., ZHANG, C., DEWITT, D., AND NAUGHTON, J. 1999. Relational databases
for querying XML documents: Limitations and opportunities. In Proceedings of the 25th International
Conference on Very Large Data Bases (VLDB). 302–314.
SHASHA, D. AND ZHANG, K. 1995. Approximate tree pattern matching. In Pattern Matching in Strings, Trees,
and Arrays. Oxford University Press.
SINGHAL, A. 2001. Modern information retrieval: A brief overview. IEEE Data Eng. Bull. 24, 4, 35–43.
SOMERVUO, P. AND KOHONEN, T. 2000. Clustering and visualization of large protein sequence databases by
means of an extension of the self-organizing map. In Proceedings of the 3rd International Conference on
Discovery Science. 76–85.
SRIKANT, R. AND AGRAWAL, R. 1996. Mining sequential patterns: Generalizations and performance improvements. In Proceedings of the 5th International Conference on Extending Database Technology (EDBT).
3–17.



25:41

TAGARELLI, A. AND GRECO, S. 2006. Toward semantic XML clustering. In Proceedings of the 6th SIAM International Conference on Data Mining (SDM). 188–199.
TAI, K.-C. 1979. The tree-to-tree correction problem. J. ACM 26, 3, 422–433.
TAMAYO, P., SLONIM, D., MESIROV, J., ZHU, Q., KITAREEWAN, S., DMITROVSKY, E., LANDER, E. S., AND GOLUB, T. R.
1999. Interpreting patterns of gene expression with self-organizing maps: Methods and application to
hematopoietic differentiation. Proc. Natl. Acad. Sci. USA 96, 6, 2907–2912.
TEKLI, J., CHBEIR, R., AND YETONGNON, K. 2009. An overview on XML similarity: background, current trends
and future directions. Comput. Sci. Rev. 3, 3.
TEKLI, J., CHBEIR, R., AND YTONGNON, K. 2007. Structural similarity evaluation between XML documents and
DTDs. In Proceedings of the 8th International Conference on Web Information Systems Engineering
(WISE). 196–211.
TRAN, T., NAYAK, R., AND BRUZA, P. 2008. Combining structure and content similarities for XML document
clustering. In Proceedings of the 7th Australasian Data Mining Conference (AusDM). 219–226.
VAKALI, A., POKORN, J., AND DALAMAGAS, T. 2004. An overview of web data clustering practices. In Proceedings
of the International Conference on Extending Database Technology Workshops (EDBI). 597–606.
VIYANON, W., MADRIA, S. K., AND BHOWMICK, S. S. 2008. XML data integration based on content and structure
similarity using keys. In Proceedings of the On the Move to Meaningful Internet Systems Conferences (1).
484–493.
VUTUKURU, V., PASUPULETI, K., KHARE, A., AND GARG, A. 2002. Conceptemy: An issue in XML information
retrieval. In Proceedings of the International World Wide Web Conference (WWW).
WANG, G., SUN, B., LV, J., AND YU, G. 2004. RPE query processing and optimization techniques for XML
databases. J. Comput. Sci. Technol. 19, 2, 224–237.
WANG, J., ZHANG, Y., ZHOU, L., KARYPIS, G., AND AGGARWAL, C. C. 2009. CONTOUR: an efficient algorithm for
discovering discriminating subsequences. Data. Min. Knowl. Disc. 18, 1–29.
WILDE, E. AND GLUSHKO, R. J. 2008. XML fever. Comm. ACM 51, 7, 40–46.
XU, R. AND WUNSCH, D. 2005. Survey of clustering algorithms. IEEE Trans. Neural Netw. 16, 3, 645–678.
YANG, J., CHEUNG, W. K., AND CHEN, X. 2009. Learning element similarity matrix for semi-structured document
analysis. Knowl. Inform. Syst. 19, 1, 53–78.
YANG, J., WANG, H., WANG, W., AND PHILIP, S. 2005a. An improved biclustering method for analyzing gene
expression profiles. Artifi. Intell. Tools 14, 5, 771–790.
YANG, R., KALNIS, P., AND TUNG, A. K. H. 2005b. Similarity evaluation on tree-structured data. In Proceedings
of the ACM SIGMOD International Conference on Management of Data (SIGMOD). 754–765.
YOON, J. P., RAGHAVAN, V., CHAKILAM, V., AND KERSCHBERG, L. 2001. Bitcube: A three-dimensional bitmap indexing
for XML documents. J. Intell. Inform. Syst. 17, 2-3, 241–254.
ZHANG, K. AND SHASHA, D. 1989. Simple fast algorithms for the editing distance between trees and related
problems. SIAM J. Comput. 18, 6, 1245–1262.
ZHANG, T., RAMAKRISHNAN, R., AND LIVNY, M. 1996. BIRCH: An efficient data clustering method for very
large databases. In Proceedings of the ACM SIGMOD International Conference on Management of Data
(SIGMOD). 103–114.
ZHAO, Y. AND KARYPIS, G. 2002a. Criterion functions for document clustering: Experiments and analysis. Tech.
rep. 01-40, Department of Computer Science, University of Minnesota.
ZHAO, Y. AND KARYPIS, G. 2002b. Evaluation of hierarchical clustering algorithms for document datasets. In
Proceedings of the ACM International Conference on Information and Knowledge Management (CIKM).
515–524.
Received March 2009; revised August 2009; accepted October 2009


Xml data clustering an overview

More Related Content

What's hot (18)

Similar to Xml data clustering an overview (20)

More from unyil96 (20)

Recently uploaded (20)

Xml data clustering an overview