Recommendation generation by integrating sequential

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Issue: 01 | Jan-2014, Available @ http://www.ijret.org 201
RECOMMENDATION GENERATION BY INTEGRATING SEQUENTIAL
PATTERN MINING AND SEMANTICS
Geethapriya Uvaraja
Computer Science and Engineering Anna University, Tamilnadu, India
Abstract
As the Internet usage keeps increasing, the number of web sites and hence the number of web pages also keeps increasing. A
recommendation system can be used to provide personalized web service by suggesting the pages that are likely to be accessed in
future. Most of the recommendation systems are based on association rule mining or based on keywords. Using the association rule
mining the prediction rate is less as it doesn’t take into account the order of access of the web pages by the users. The
recommendation systems that are key-word based provides lesser relevant results. This paper proposes a recommendation system that
uses the advantages of sequential pattern mining and semantics over the association rule mining and keyword based systems
respectively.
Keywords: Sequential Pattern Mining, Taxonomy, Apriori-All, CS-Mine, Semantic, Clustering
----------------------------------------------------------------------***------------------------------------------------------------------------
1. INTRODUCTION
A recommender system understands the users’ navigation
pattern by exploiting web usage mining and provides
personalization service based on the results of mining. The
recommender system will propose links of possible interest to
the user. Data mining and machine learning algorithms are
used in developing a recommendation system. In general the
recommendation systems take the users’ navigations in the
form of log file. Some kind of mining is performed over the
data in the log file to find the usage patterns. When a new user
arrives, the current access is matched with the patterns that are
mined to generate recommendations.
The semantic information is used along with web usage data
to get more relevant patterns. Since web logs lack semantic
information about web pages visited by users, it is difficult to
understand the preferences and intents of users. With the
development of the Semantic Web, semantics in web content
can be used for improving the relevancy of recommendation.
The objective of this paper is to provide recommendations by
integrating sequential pattern mining and semantics. The
sequential patterns obtained from two sequential pattern
mining algorithms Apriori-All and Conditional Sequence
Mine are compared. Taxonomy is constructed for the website
for which the recommendations are to be generated. The
keywords representing each web page of the website are found
and they are mapped with the taxonomy to obtain the
categories by using the similarity measure. The documents are
then clustered based on the categories. The recommendation
rules generated by sequential pattern mining algorithm and the
clusters are used to generate recommendations.
The rest of the paper is organized as follows. Section 2
describes the literature survey and the related work. Section 3
presents the system design. Section 4 describes the results of
the modules implemented and performance evaluation.
Section 5 is about the conclusion.
2. LITERATURE SURVEY AND RELATED
WORK
Web mining is the process of mining or extracting useful or
valuable information on web. Web mining is an application of
data mining techniques on World Wide Web to extract
patterns from resources available on web. Web mining has
three categories. They are
1. Web content mining
2. Web usage mining
3. Web structure mining
Web content mining is the discovery of patterns from contents
of the web pages for improving the relevancy in the field of
web mining. It involves information extraction, knowledge
discovery and analysis of a collection of documents. Web
usage mining is the discovery of the patterns from user
navigation data in the form of log file. Web structure mining is
the discovery of knowledge from hyperlinks and link structure
between the web pages.
A recommendation system incorporates the web mining
techniques in order to provide personalized suggestions to the
users. There are various approaches through which
recommendations can be generated. Most of the traditional
recommendation systems are based on clustering, the user
given rating and feedback provided by the users. Due to the

__________________________________________________________________________________________
presence of a huge number of meaningful clusters and profiles
for visitors of a typical highly rated Website, the model-based
or distance-based method tend to make too strong and
simplistic assumptions and becomes excessively complex and
slow. Collaborative filtering systems predict a person’s
affinity for items or information by connecting that person’s
recorded interests with the recorded interests of a community
of people and sharing ratings between likeminded persons.
This approach is based on an assumption that those who
agreed in the past tend to agree again in the future. It cannot
acquire accurate recommendation results when user rating data
are extremely sparse. This approach can make suggestions to a
user that are outside the scope of previous selected items. But
it not scalable when number of users and items increases. A
content-based filtering system selects items based on the
correlation between the content of the items and the user’s
preferences as opposed to a collaborative filtering system that
chooses items based on the correlation between people with
similar preferences.
In the personalized recommendation systems, which are based
on web usage mining, association-mining technology was
applied to predict user-browsing behavior. This method scales
better with large datasets compared to systems based on
clustering. Weighted association rule model was proposed by
assigning a significant weight to the pages based on time spent
by each user on each page and visiting frequency of each
page. The weighting measure was used to judge the
importance of a page to a user, and try to give more
consideration to pages which are more useful to the user. This
method is better in precision and coverage rates than the
conventional association rule based recommendation [8]. The
drawback with this approach is that the idle time of the user on
a web page cannot be differentiated from the actual time spent
by the user on a page. A recommendation method was
proposed to be applied to web log mining by integrating user
clustering and association rule mining techniques. The
precision by this method was better compared to association
rule based recommendation system. But there is no much
improvement in coverage ratio and speed [7].
Semantic Web Mining aims at combining the two fast-
developing research areas Semantic Web and Web Mining.
Semantic web makes the web contents understandable not
only by human but also by the machine. Machine processable
information can provide more relevant results and can
improve the precision and recall. To achieve this, the content
of the semantic web are mostly represented by ontology,
XML, RDF and meta-data. Ontology Learning is a method for
extraction of semantics from the Web in order to create
ontology. Machine learning techniques were used to improve
the ontology engineering process. Semantic Web Mining
improves the results of Web Mining by exploiting the new
semantic structures in the web [10].
3. SYSTEM DESIGN
The overall system architecture of the proposed
recommendation system is the Fig. 1.
Fig. 1 System Design
System Description
The dataset used is the web server log file (web usage data) of
the website http://cs.annauniv.edu and the web pages (web
content data) of the same site. Each module in the designed
recommendation system is explained as follows.
3.1 Pre-processing
The objective of pre-processing is to reduce the size of the
web usage data (web server log) and to increase the quality of
the data that will be suitable for mining. Web usage data
contains lot of noise. The log file contains records for the
requests to images, multimedia files or script files. Some
records are due to the requests made by the web crawlers. Log
file also contain the requests that are not processed. Those
records are not necessary for analysis and hence they are
eliminated. The un-processed requests can be identified using
the http status code. If the requested page has images many
records are created for a single web page request. Such
repetitive records are pruned. A web server would receive
request from multiple users at same time. The records for the
request from different users will be interleaved. The session
for each user has to be identified. A user can be identified as a
combination of the IP address and the agent. The list of web
pages requested by a user forms a web access list which is
input to the sequential pattern mining algorithms.
3.2 Sequential Pattern Mining
The sequential pattern mining is the mining of frequently
occurring patterns ordered by time. Using sequential pattern
mining, one can identify the paths that users frequently follow

__________________________________________________________________________________________
on a web site and hence it increases the prediction rate.
Sequential pattern mining is well suited for log study due to
the sequential nature of web users’ activity. In this paper the
frequent sequence patterns generated by two sequential pattern
mining algorithms AprioriAll and CS-Mine are evaluated.
Apriori-All is similar to the well known Apriori algorithm but
the difference being that the Apriori-All considers the ordering
of the items in a transaction where as Apriori doesn’t take into
account the order of items in a transaction. Hence Apriori-All
can be said as the sequential version of Apriori [3].
Conditional Sequence Mine algorithm works on the
conditional sequence of each frequent pattern given a
sequence database where each sequence is a list of
transactions ordered by transaction time as an input to the
algorithm [1].
3.3 Pattern Tree Construction
A pattern tree is used to store the sequential patterns
compactly. A single scan of sequential access patterns
generated by the AprioriAll or Conditional Sequence mine
algorithms is necessary to construct a pattern tree. A pattern
tree is based on trie data structure. Trie is used for storing
strings to enable fast pattern matching. The root node of the
pattern tree is a dummy node. All other nodes have a web
access and its support. All the sequential access patterns in the
pattern tree can be visited by following the path starting from
the root node of the tree.
3.4 Sequence Pattern Based Recommendation Engine
This component searches for the best matching access path in
the Pattern-tree for the given access sequence. The given
access sequence is matched with the nodes in the pattern tree
and the web accesses in the child nodes are generated as
recommendations. The recommendations generated are in the
descending order of the support of the web accesses i.e., the
most frequent access is suggested first.
The suffix sequences of the current access sequence will be
considered when the matching path of the whole access
sequence cannot be found. Search will be performed on the
matching path based on the same access sequence by
removing the first item repeatedly until a matching path is
found or when no more items can be removed from the access
sequence. The length of the longest path in the Pattern-tree is
the depth of the Pattern-tree. The matching path will not exist
when the length of the current access sequence is longer than
the depth of the Pattern-tree. Therefore, some initial items can
be removed to make the current access sequence shorter than
the depth of the Pattern-tree before the sequence matching
process.
The recommendation rules generated by this engine are later
used by semantic based recommendation engine to generate
recommendations based on semantics of the web contents.
3.5 Keyword Extraction
The web pages are parsed, the tags are removed and the web
contents are extracted. A text document is created
corresponding to each page. Most common words (stop
words) are removed from each document. The significant
keywords for each document are identified using TF*IDF.
TF*IDF is a statistical measure used to evaluate how
important a word is to a document in a collection. 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.
3.6 Taxonomy Construction
The taxonomy is a concept hierarchy that provides a means for
designing enhanced searching, browsing and filtering systems.
A domain specific taxonomy is constructed using XML. The
keywords extracted in the previous step have to be mapped
with the elements in the taxonomy to get the categories. Hence
the taxonomy has a greater influence on the outcome of the
mapping process. For this purpose the taxonomy has to be
semantically related to the contents of the website.
3.7 Mapping Keywords and the Categories
Keywords are the representatives of the contents of the web
pages. These keywords are mapped with the categories in the
taxonomy using the thesaurus (Wordnet). If a keyword exists
in the taxonomy, it is included. If it doesn’t exist, a closest
category in the taxonomy is found by making use of the
thesaurus. In this paper the closest category in the taxonomy is
found using Jiang and Conrath similarity measure. Now the
documents are represented as the categories in the taxonomy.
3.8 Document Clustering
As the web pages are now represented as the categories in the
taxonomy, they are clustered based on the similarity between
the categories in the taxonomy. Clustering based semantics
would aid the recommendation generation easier. DBSCAN is
a density based clustering algorithm used for document
clustering.
3.9 Semantic Based Recommendation Engine
The recommendation rules generated by the sequential pattern
based recommendation engine and the document clusters are
provided as input to the semantic based recommendation
engine. Every document is already assigned to a relevant
cluster. Given a recommendation rule, all the URLs in the
cluster in which the URL in the right hand side of the rule
belongs are generated as recommendations [10].
4. RESULTS
Data set used is the log file of the website
http://cs.annauniv.edu. The log file is assumed to be static for

__________________________________________________________________________________________
experimental purpose. Number of records in the log file is
140283. After preprocessing the size of the records were
reduced to 15% of the initial size. Web access lists are
extracted from the preprocessed records. The web access lists
generated are used by the sequential pattern mining algorithms
Apriori-All and CS-Mine and their performance is evaluated.
Fig. 2 Run time vs. Threshold Support
Fig 2 shows that the run time of CS-Mine is lesser compared
to AprioriAll. This is because AprioriAll generates huge set of
candidate sequences and needs many scans of the sequence
database. After a certain support threshold, say 6% the run
time of both the algorithms becomes equal. The run time of
both the algorithms decreases as the support threshold
increases.
Fig. 3 Number of patterns generated vs. threshold support
Fig 3 shows that the number of patterns generated by
AprioriAll is greater than that of the CS-Mine algorithm.
Though the number of patterns generated by Apriori-All was
greater than that of the CS-Mine, recommendation engine
generated same recommendations using both the algorithms at
all the threshold support. Hence CS-Mine which is lesser time
consuming by generating fewer patterns is found to be
efficient in terms of recommendation generation also.
Precision is an evaluation measure used to find how probable
a user will access one of the recommended pages.
Precision = Number of correct recommendations / Total
number of recommendations
Fig. 4 Precision percentage vs. No. of pages recommended
Fig 4 shows that the precision of the recommendation system
increases as the number of the recommendations generated
increases. Since many recommendations will be generated at
lower support threshold, the precision will be high at low
threshold support.
CONCLUSIONS
In this paper the recommendations are provided by exploiting
two sequential pattern mining algorithms – AprioriAll and CS-
Mine and it has been shown that the CS-Mine algorithm is
efficient in terms of time, generating less number of patterns.
The recommendation obtained using both the algorithms are
same. Rules generated by the sequential pattern based engine
are then used to provide semantic based recommendations.
Sequential pattern based recommendations are generated by
matching the current access of a user with paths in the pattern
tree. During the pattern matching, only few paths in the
pattern tree are matched. Hence there is a reduction in time as
not all the paths in the tree are traversed. Also there is a
reduction in space as all the sequences are compacted in this
pattern tree and hence the sequences with same prefixes share
common sub paths. The time consuming part of this approach
is the pattern tree construction. In real time the pattern tree
construction need not be performed for every request and
hence it doesn’t greatly affect the time when providing
suggestions. One of the challenging problems in this
recommendation system is when a page is newly visited and is
not in the pattern tree, no recommendations are provided
initially. When the pattern tree is updated with the new access
sequences then the recommendations will be generated.

__________________________________________________________________________________________
REFERENCES
[1]. Xiaogang Wang; Yan Bai; Yue Li: “An Information
Retrieval Method Based On Sequential Access Patterns”
Wearable Computing Systems (APWCS), 2010 Asia-Pacific
Conference, April 2010, pp: 247 - 250
[2]. Jaideep Srivastava, Robert Cooley, Mukund Deshpande,
Pang-Ning Tan: “Web Usage Mining: Discovery and
Applications of Usage Patterns from Web Data” ACM
SIGKDD Volume 1, Issue 2, January 2000, pp: 12 – 23
[3]. WANG Tong, HE Pi-lian: “Web Log Mining by an
Improved AprioriAll Algorithm” World Academy of Science,
Engineering and Technology, April 2005
[4]. Baoyao Zhou, Siu Cheung Hui, Kuiyu Chang: “An
Intelligent Recommender System using Sequential Web
access patterns” Cybernetics and Intelligent Systems, 2004
IEEE Conference, 2004, pp: 393-398
[5]. Yi Dong, Huiying Zhang, Linnan Jiao: “Research on
Application of User Navigation Pattern mining
recommendation” Intelligent Control and Automation, 2006.
The Sixth World Congress, 2006, pp: 6106 - 6110
[6]. Forsati, R: Meybodi, M.R.; Rahbar, A: “An Efficient
Algorithm for Web Recommendation Systems” Computer
Systems and Applications, 2009.IEEE/ACS International
Conference, 2009, pp: 579 - 586
[7]. Jaideep Srivastava , Robert Cooleyz , Mukund
Deshpande, Pang-Ning Tan: “Web Usage Mining: Discovery
and Applications of Usage patterns from web data”
ACM SIGKDD Volume 1, Issue 2, Jan 2000, pp: 12-23
[8]. Bettina Berendt, Andreas Hotho, and Gerd Stumme:
“Towards Semantic Web Mining”, Horrocks and J. Hendler
(Eds.): ISWC 2002, LNCS 2342, 2002, pp. 264–278
[9]. Lappas, G: “An Overview of Web Mining in Societal
Benefit Areas” 4th IEEE International Conference on
Enterprise Computing, E-Commerce and E-Services, 2007,
pp: 683-690
[10]. M. Eirinaki, M. Vazirgiannis, I. Varlamis: “SEWeP:
Using Site Semantics and a Taxonomy to Enhance the Web
Personalization Process” SIGKDD ’03, August 24-27, 2003
BIOGRAPHIE:
Geethapriya Uvaraja, Post-Graduate from Anna
University, Chennai

Recommendation generation by integrating sequential

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Recommendation generation by integrating sequential (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

Recommendation generation by integrating sequential