RECOMMENDATION FOR WEB SERVICE COMPOSITION BY MINING USAGE LOGS

International Journal of Data Mining & Knowledge Management Process (IJDKP) Vol.6, No.2, March 2016
DOI : 10.5121/ijdkp.2016.6207 83
RECOMMENDATION FOR WEB SERVICE
COMPOSITION BY MINING USAGE LOGS
Vivek R, Prasad Mirje and Sushmitha N
Department of Information Science, RVCE, Bangalore, India
ABSTRACT
Web service composition has been one of the most researched topics of the past decade. Novel methods of
web service composition are being proposed in the literature include Semantics-based composition, WSDL-
based composition. Although these methods provide promising results for composition, search and
discovery of web service based on QoS parameter of network and semantics or ontology associated with
WSDL, they do not address composition based on usage of web service. Web Service usage logs capture
time series data of web service invocation by business objects, which innately captures patterns or
workflows associated with business operations. Web service composition based on such patterns and
workflows can greatly streamline the business operations. In this research work, we try to explore and
implement methods of mining web service usage logs. Main objectives include Identifying usage
association of services. Linking one service invocation with other, Evaluation of the causal relationship
between associations of services.
KEYWORDS
Web service composition, Recommender system, Multi level Association rule, Data mining, Web Service
invocation logs
1. INTRODUCTION
Nowadays a significant part of a software system is structured using software services and
implemented using Web service technologies. Web services have become standard, collaborative
and can handle a complex conversation with other software and service systems. Service
composition is implementation technology that defines how to implement a web service by
composing other services. Web service composition is essentially operation composition provided
by web service. Business process architects discover [10] required web service to achieve a given
goal and compose them into an application using specification language or service composition
language (e.g., XL, WSFL, and BPML [10]). The main challenges in service composition are a
lack of domain knowledge about a business process which is being implemented and identify a
set of service that can be used to execute a business process [11].
To address aforementioned challenges many types of researches have been carried out, they are
divided into two paradigms, one where we define business process first (goal achieved by
composed service) and then identify set of service which satisfies the business process, this can
paradigm can be regarded as business process oriented service composition (BPOSC). The other
approach would be to identify set of service which are frequently being invoked, associate them
and identify a workflow from the association and evaluation of workflow in terms of validity and
commercial benefits, such composition paradigm can be regarded as invocation or Execution
based service composition (EBSC). Recommendation systems for BPOSC aid the user in
identification of business process, defining the business process and searching of services which
satisfies the business process, to this regard many types of research have been carried out, QoS
based searching [9], and syntactic & semantic based searching methods have been developed.

84
Server logs are rich input to provide recommendations for EBSC, many data mining techniques
like association rule mining, clustering can be easily adopted which can aid the user in associating
web services, identifying workflows and validation of workflows
In this research work, we try to explore and implement methods of mining web service invocation
logs. Main objectives include Identifying usage association of services. Linking one service
invocation with other. We will try to evaluate popular data mining techniques which can be used
to discover association among web service invocation and visualize these association as a
recommendation for service composition.
The remainder of this paper is organized as follows. Section 2 discusses the related work. Section
3 gives an overview of the dataset required for such recommendation and challenges and
approach followed in generation of same. Section 4 evaluates the various association mining
algorithm and gives details of our approach. Section 5 discusses the results of the approach
followed; Section 6 concludes the paper.
2. RELATED WORK
There are many approaches to service composition and each approach addresses a different set of
challenges of service composition. These approaches can be classified into four different groups
[12] process based, model based, logic or calculi based, and AI-based approaches. As described
in the previous section this classification applies to business process oriented service composition.
Our work presented here relates to execution based service composition approaches. The concept
of application of data mining is explored in [4], recommends a set of applications that can
leverage problems concerned with the planning, development and maintenance of Web services
and their compositions. [14] Presents a similar approach of applying Web service log-based
analysis and process mining techniques in order to provide a mean to ensure a correct and reliable
Web services modelling and execution. They also present the issues and challenges in collecting
execution logs of service invocation. [13] Proposes an approach that automatically identifies the
service composition patterns from various applications using execution logs. The approach
employs Apriori algorithm to mine associated services. Analysing the control flow of services
and the order of service invocation based on the execution logs. In our research we go beyond the
use of simple Apriori for associating web services, we exploit the natural hierarchy that exists
between service and operation, in order to generate better and reliable service association that can
be used as a recommendation for service composition.
3. DATASET GENERATION
Although weblogs mining is an emerging research topic for recommendation systems, the
publically available datasets for researchers are more concentrated towards the recording of
invocation of web pages rather than web service. Such weblogs record the hits a web page has
received. Available datasets for research under enhancement of web services are either QoS based
[6, 7, 8] or WSDL-based [9]. Minimalistic web service invocation logs required should contain
parameters like time of invocation, client IP or session identifier and web service URL,
additionally the logs may also contain QoS information like data size, response time, HTTP code
and message.
Due to lack of such detailed web service invocation logs, we have resorted to the generation of
web service logs. This section gives brief explanation of methods followed during data set
generation. The generated dataset consists of parameters as shown in Table 3, for simplicity we
have considered web service name instead of web service URL. Usually, a single web service
exposes multiple operations, which is indicated by operation column. In a composite web service,

85
the order in which these web services are invoked are also important, this order of invocation is
governed by WS coordination protocol. The recommender system should maintain this sequence
of invocation.
Table 1: Generated dataset
Session
id
Time stamp Web service
name
Operation Response time Response size
... … … … …. ….
We consider a set of 50 compositions consisting of 100 web service (WS1, WS2…WS100), each
web service exposes randomly chosen number of operations ranging from 2 to 15 as our final
result to be produced. Using the same 100 web services, we generate invocation logs, the
generated logs consist of 1000 session. In each session, a randomly chosen web service will be
invoked. The order of invocation is also random. One may argue that the occurrence of any of the
composition in the above-said list would be very sparse due to randomness involved in the
generation of logs. When any mining technique applied on such sparse data, a threshold for
interestingness criteria (like support and confidence under association rules) which would help us
recommend service composition should be set to a lower value. This would lead in increased
number of candidate compositions, and also reduced the performance of the mining algorithm.
Hence, for every randomly chosen invocation between range of 2 to 10, one of the compositions
in the list will be randomly chosen and included as subsequence in the invocation logs. This
would bring a balance between randomness and sparse-city in the data and generate fair web
service invocation logs, which can be used to evaluate different mining procedure best suited for
recommender system of WS compositions.
4. PROPOSED METHODOLOGY
Table 2: Performance of various algorithm
Algorithm Min
support
(%)
Min
confidence
(%)
Candidate
sequence
count
Frequent
sequence
count
Number of
association rules
Apriori 3 3 149490 115001 25366354
Apriori 3.5 3.5 53038 49624 12664936
Apriori 3.5 3.8 49624 49624 12664936
Apriori 3.7 3.7 87 62 360
Apriori 3.8 3.8 63 63 602
Closed
association
rules
3.7 3.7 2 2 60
FP growth 3.7 3.7 62 62 360
Table 2: shows performance of various well-known data mining techniques that can be used for
generation of associations of web composition. A set of 1000 sessions consisting of service
invocation was used as input. We can observe that when the overall value of mini support & min
confidence is low, and chosen different value either generates too many candidates or too few of
them, which in turn results in varying number of association rules. We can see that the efficiency
of such algorithm greatly depends on the interestingness criteria given as input (support and
confidence).
The observed efficiency can be greatly increased by modelling some application specific
knowledge into the system. Service specific information like description of service [1] in WSDL,
QoS [6, 7] provided by service, invocation order provided coordination protocol can be used. Use

86
of service description and QoS parameter have been greatly explored in the literature. In this
paper, we will explore the invocation order as a parameter that can benefit the recommender
system. Invocation order can either be derived from coordination protocol or discovered from
invocation logs. Invocation order derived from coordination protocols provide as information
regarding service composition in practice, they help in verifying whether implemented service
compositions are of value addition or not, rather than helping to discover new service
composition. On the other hand discovered order can be used to identify new service
composition.
Figure 1: Concept hierarchy for multilevel association rule mining
Figure 1 shows the hierarchal view of the composed list of service, i.e., each web service
performs or consists set of operations. The findings of service composition can be done iteratively
or step-wise using this hierarchy. This technique has an intrinsic way to reduce candidate list.
Learning of one iteration is used to reduce the candidate set to be processed in subsequent phases.
Figure 2: Proposed methodology for web service composition using multi-level association rules
Figure 2 gives the overview of the approach to extract service composition recommendation from
execution log. Invocation logs captured by the web server is given as input to the system. Logs
are pre-processed as initial step which includes process of identification of sessions, client ids,
client groups and transformations making the logs minable by data mining methods. In next step
the whole logs are scanned and from identified sessions, web service names are extracted
excluding the operations involved in the invocation. If a web service is invoked multiple times
with same operations (as in loop invocation) or different operations, single occurrence will be
considered, for e.g. consider a session with invocation {A (a), A (b), C (d), D (a), A (c)} (where
A, C, D are web service names and a, b, c, d are operations with-in web service) will be
transformed to {A, C, D}. Such transformations which contain only web service names forms
level one of the concept hierarchy. To preserve the ordering in composed service, invocations in a
session are treated as sequence. These candidate sequences can occur as complete sequence or
sub-sequence in sessions, the number of occurrence in logs forms the support of candidate
sequence. There are many ways to find interestingness of a candidate, support being most widely
used, we take average support as the threshold. Each sequence corresponds one or more sessions,

87
only those sessions with candidate sequence or sub-sequence whose support is greater than avg.
support will be considered for next level of mining. This greatly reduces the number of candidates
to be processed, enhances speed and accuracy of the system. Reduced invocation logs will be
generated which will serve as input to level two candidate web service generation. The frequent
sequence will be found consisting of service names as well as operation being invoked. These
candidate WS sequence can be used for generation of association rules, top n rules will be
provided as the recommendation to the user.
5. EXPERIMENTS AND RESULTS
As discussed in previous sections, a data set of 1000 session was used as input for generation
service composition. As discussed in section 3, the evaluating criteria considered in this work is a
set of 50 compositions consisting of 100 web services. Association rules generated from proposed
method are compared against these to determine the performance.
Table 3: Experimental results
Algorithm Min
support
(%)
Min
confidence
(%)
Candidate
sequence
count
Frequent
sequence
count
Number of
association
rules
Number of
matching
association
Apriori 3.5 3.5 53038 49624 12664936 50
Apriori 3.5 3.8 49624 49624 12664936 50
Apriori 3.7 3.7 87 62 360 16
Proposed
Method
3.5 3.5 899 530 1370 48
Proposed
Method
4.5 4.5 121 188 156 42
Proposed
Method
6.5 6.5 84 135 101 35
The correctness of approach proposed can be verified by the use of two metrics [13], precision
and Noise ratio. Precision is given by equation 1; Precision obtained should be high, as it measure
whether the proposed approach is able to identify the associations present in the dataset
(execution logs). Here precision can also be considered as accuracy.
Noise vs expected association rules ratio is given by equation 2; Noise ratio should be low, as it is
the measure of ability of proposed approach to identify interesting composition from given
execution logs. It also signifies the capacity of proposed approach to reduce the candidate set into
an association rule. Ideally the ratio obtained should be 1, i.e., when expected number of
association is equal to generated number of association. Less ratio signifies possible loss of data
or over reduction, and more ratio signifies introduction of noise in the association rules generated.
Noisy set of association rules would not be beneficial for recommender system.
Table 3 shows the experimental results in identifying the service composition from invocation
logs. Although the performance of Apriori is more accurate when compared to proposed method,
the ratio of noise to expected association rules is significantly high, making Apriori
computationally heavy, whereas the proposed method would generated almost accurate result

88
with very much less noise in generated association rules. Figure 3 shows the graphical form of
Noise vs Expected association rules. Table 4 provides the results obtained from calculations of
the metrics. For calculation, the number of expected association is considered as 50, as it the
number of composition with which the data set was generated.
Table 4: Correctness metrics for proposed method
Algorithm Number of association
rules
Number of matching
association
Precision
(%)
Noise ratio
Apriori 12664936 50 100 253298.72
Apriori 12664936 50 100 253298.72
Apriori 360 16 32 7.2
Proposed
Method
1370 48 96 27.4
Proposed
Method
156 42 84 3.12
Proposed
Method
101 35 70 2.02
6. CONCLUSION
In this research work, a novel approach for extraction of recommendations for web service
composition is proposed. The proposed method employs the intrinsic hierarchy available in web
service (i.e., every service is composed of set of operations) to reduce the candidate service
sequence to be processed. This yields in associations which has higher support. The proposed
work also explores invocation order of web services as criteria for service composition
recommendation. From the results shown we can conclude that proposed method would generated
much lesser noise association rules, hence providing more interesting recommendation for web
service compositions.
Figure 3: Noise vs Expected Association rules
REFERENCES
[1] Liang Chen, Jian Wu, Zibin Zheng, Michael R. Lyu, Zhaohui Wu, Modeling and exploiting tag
relevance for Web service mining, Knowledge and Information Systems, April 2014, Volume 39,
Issue 1, pp 153-173

89
[2] Richi Nayak, Aishwarya Bose, A Data Mining Based Method for Discovery of Web Services and
their Compositions, Real World Data Mining Applications, Annals of Information Systems Volume
17, 2015, pp 325-342
[3] Nayak, Richi (2008) “Data mining in web services discovery and monitoring”, International Journal
of Web Services Research, 5(1). pp. 62-80.
[4] Nayak, Richi and Tong, Cindy M. (2004) “Applications of data mining in web services”, In
Proceedings 5th International Conferences on Web Information Systems, Brisbane, Australia
[5] Bhuvaneswari V and Umajothy P, “Mining Temporal Association Rules from Time Series Microarray
Using Apriori Algorithm”. In proceedings of Review of Bioinformatics and Biometrics (RBB)
Volume 2 Issue 2, June 2013
[6] Zibin Zheng, Michael R. Lyu, "Collaborative Reliability Prediction for Service-Oriented Systems", in
Proceedings of the ACM/IEEE 32nd International Conference on Software Engineering (ICSE2010),
Cape Town, South Africa, May 2-8, 2010, pp. 35 – 44.
[7] Zibin Zheng, Yilei Zhang, and Michael R. Lyu, “Distributed QoS Evaluation for Real-World Web
Services,” in Proceedings of the 8th International Conference on Web Services (ICWS2010), Miami,
Florida, USA, July 5-10, 2010, pp.83-90.
[8] Yilei Zhang, Zibin Zheng, and Michael R. Lyu, “Exploring Latent Features for Memory-Based QoS
Prediction in Cloud Computing,” in Proceedings of the 30th IEEE Symposium on Reliable
Distributed Systems (SRDS 2011), Madrid, Spain, Oct.4-7, 2011.
[9] Yilei Zhang, Zibin Zheng, and Michael R. Lyu, "WSExpress: A QoS-aware Search Engine for Web
Services," in Proceedings of the 8th International Conference on Web Services (ICWS2010), Miami,
Florida, USA, July 5-10, 2010, pp.83-90
[10] J. Lu, Y. Yu, D. Roy, and D. Saha, “Web Service Composition: A Reality Check”, 8th International
Conference on Web Information Systems Engineering, Nancy, France, December 3-7, 2007
[11] Bipin Upadhyaya, Ying Zou1, Shaohua Wang and Joanna Ng, “Automatically Composing Services
by Mining Process Knowledge from the Web”
[12] Baryannis, George, and Dimitris Plexousakis. "Automated Web Service Composition: State of the Art
and Research Challenges." ICS-FORTH, Tech. Rep 409 (2010).
[13] Tang, Ran, and Ying Zou. "An approach for mining web service composition patterns from execution
logs." Web Systems Evolution (WSE), 2010 12th IEEE International Symposium on. IEEE, 2010.
[14] Gaaloul, Walid, Karim Baïna, and Claude Godart. "Log-based mining techniques applied to web
service composition reengineering." Service Oriented Computing and Applications 2.2-3 (2008): 93-
110

RECOMMENDATION FOR WEB SERVICE COMPOSITION BY MINING USAGE LOGS

More Related Content

Viewers also liked (20)

Similar to RECOMMENDATION FOR WEB SERVICE COMPOSITION BY MINING USAGE LOGS (20)

Recently uploaded (20)

RECOMMENDATION FOR WEB SERVICE COMPOSITION BY MINING USAGE LOGS