Evaluation of models for predicting user’s next request in web usage mining

International Journal on Cybernetics & Informatics (IJCI) Vol. 4, No. 1, February 201
DOI: 10.5121/ijci.2015.4101 1
Evaluation of Models for Predicting User’s Next
Request in Web Usage Mining
Bhawna Nigam
Department of Information Technology ,Institute of Engineering & Technology,
DAVV Indore, M.P., India
Dr. Sanjiv Tokekar
Department of Electronics & Telecommunication Engineering Institute of Engineering &
Technology, DAVV Indore, M.P., India
Dr. Suresh Jain
Department of Computer Engg. Prestige Institute of Engineering Management &
Research, Indore, INDIA
Abstract
Prediction of web user behavior is the demand of today competitive edge of World Wide Web. Predicting the
next web page is not sufficient, evaluation of prediction models is important because every model have its
own pros and cons. Prediction results will be helpful if high prediction accuracy is achieved with minimum
complexity, which are depended on the prediction model. Various models and their variations are proposed
for predicting the next web page accessed by the web user. Markov model and their variations are found
suitable for web prediction. In this research we have evaluated and compared various models for predicting
next web page accessed by the web user. Experiments are conducted on three different real datasets.
Keywords
Web usage mining, Markov Model, User Navigation Session, web log and web prediction
1.Introduction
Web prediction is a field of web usage mining in which next accessed web page by web user is
predicted. Prediction results can be used for personalization of web, reducing the server
response time with proper prefetching and caching strategies [1]. It can provide guidelines for
improving the design of web applications, e-commerce to handle business specific issues like
customer attraction, customer retention, cross sales and customer departure.
Prediction of next access web page can be achieved through modelling the web log with the help
of model. The logging information is stored in a file known as web log file which resides on web
server, proxy server or client cite. The web log file is the text file which contains lots of
information such as IP address, date, time, request type etc., so it is preprocessed before
modelling. From the preprocessed web log information the user navigation session prepared. The
user navigation session is finally modeled through a model. Once the user navigation model is
ready, the mining task can be performed to predict next accessed web page. In prediction model
log file is divided into two parts training file and testing file, training file is used to build the
model and testing file is used to test the model. Various models have been proposed to
accomplish this task such as Markov Model, Semantic Model, Dynamic Nested Markov Model,
association rule mining and many more[3,4].

2
Predicting the next web page is not sufficient, it is also very important to evaluate the prediction
because every model has their advantages and limitations [9]. Lower order Markov model is less
complex with low accuracy and high order Markov model has high accuracy with high
complexity and low coverage. So before uploading any prediction model on the server, it is very
necessary to find limitations of the model. Various parameters are there to evaluate the prediction
such as how much time required by model to predict next web page, prediction accuracy,
generation time, coverage and many more. In this work we have evaluated and compared
different models for mining the web log to predict next accessed web page.
2.Related Work
Several authors have proposed models for modelling the user navigation session to predict next
accessed web page. Markov model is widely used for modelling the web navigation sessions.
Markov model is based on a well established theory. F. Khalil et al. [5], have proposed a new
framework for predicting the next web page access. In this they have study the Markov model for
prediction. If the Markov model is not able to predict the next web page then the association rule
are used for predicting the next web page. They have also proposed that if there will be ambiguity
in the prediction, it will be resolve by association rule. J. Borges et al. [6, 7, 8], have proposed the
Higher-order Markov model with clustering technique to improve the effectiveness of Markov
Model. The K-mean clustering technique has been used to reduce the state space complexity. F.
Khalil et al. [10] have proposed the integrated approach for predicting the next access web page
where they have tried to achieve the high level of predictive accuracy with low state space
complexity. Siriporn Chimphlee et al. [11], used association rule for next access prediction. Nizar
R. [12] proposed semantic rich markov model for web prefetching. B. Nigam et al. [13] used the
concept of dynamic nested markov model to predict next accessed web page whose analysis is
done on different schemes of prefetching and caching [14]. M.T. Hassan et al. [17] presented
Bayesian Models for two things like learning and predicting key Web navigation patterns. Instead
of modelling the general problem of Web navigation they focus on key navigation patterns that
have practical value. Mamoun A. Awad et al. [18], analyzed and studied Markov Model with all-
Kth Markov Model for web prediction. They proposed a new modified Markov Model to
alleviate the issue of scalability in the number of paths. Poornalatha G et al. [19] presented a
paper to solve the problem of predicting the next web page to be accessed by the user based on
the mining of web server logs that maintains the information of users who access the web site.
Section 2, describes the Markov Model and Dynamic Nested Markov Model, Section 3, describes
the experimental results and finally section 4 describes the future work and conclusion.
3.Prediction Models
Markov Model is compact, simple, expressive and based on a well-established theory. Markov
Model is widely used to model user navigation sessions. In first-order Markov Model, each state
corresponds to a web page and each pair of viewed web page corresponds to state transition. Two
artificial state i.e. start and final, are incorporated in the model. In second-order Markov Model,
each state corresponds to sequence of two viewed web pages and so on.

3
A) Generation Second-Order Markov Model
Markov Model is widely used to model user navigation sessions. Two artificial states i.e. start and
final are incorporated in the model. In first-order Markov Model, each state corresponds to a web
page and each pair of viewed page corresponds to state transition. In second-order Markov
Model, each state corresponds to sequence of two viewed web pages and so on. Figure 5.2 shows
the corresponding transition diagram of Hypertext Probabilistic Grammar. Hypertext Probabilistic
Grammar (HPG) is a four-tuple <V, Σ, S, P>, V={A1, A2, A3….} is set of non-terminals , Σ ={a1,
a2, a3….} is set of terminal symbol, S is start symbol , P is set of production rule.
if the probability of state being in a start production is proportional to the total number of
times the state was visited in the collection of navigation sessions. Therefore, when the
probability of a start production is proportional to the number of times the corresponding state
was visited, implying that the destination node of a production with higher probability
corresponds to a state that was visited more often. The parameter can take any value between 0
and 1, providing a balance between the two scenarios described above. As such, gives the
analyst the ability to tune the model for the search of different types of patterns in the user
navigation. Finally, the probability of a transitive production is assigned in such a way that it is
proportional to the frequency with which the corresponding link was traversed.
Table 1 shows the example of collection of training and testing user navigation sessions. T1 to T5
are the transaction ID. There are five web pages P1 to P5.
Table 1: Collection of User Navigation Sessions.
Transaction ID Training Sessions
T1 P2, P3, P1, P5
T2 P2, P1, P3, P4, P5
T3 P1, P2, P5
T4 P1, P5, P4
T5 P1, P2, P4
Transaction ID Testing Sessions
T1 P3, P1, P4
T2 P1, P5, P4
T3 P4, P5

4
Fig. 1: Second-order Markov Model corresponds to training file of table 1.
Fig. 1 shows the second-order Markov Model corresponds to the training file of table 1. The model
is represented with the hypertext weighted Matrix. Here states are the sequence of two viewed web
page, S is the start state and F is final state.
Table 2 shows Hypertext weighted Matrix for second-order Markov Model. if the state exists then
the weight shows the count of number of times the sequence occurs in the training file otherwise
it will be 0.
Table 2: Hypertext Weighted Matrix.
2nd
Order Markov Model P1 P2 P3 P4 P5
{P1, P2} 0 0 0 1 1
{P1, P3} 0 0 0 1 0
{P1, P5} 0 0 0 1 0
{P2, P1} 0 0 1 0 0
{P2, P3} 1 0 0 0 0
{P2, P4} 0 0 0 0 0
{P2, P5} 0 0 0 0 0
{P3, P1} 0 0 0 0 1
{P3, P4} 0 0 0 0 1
{P4, P5} 0 0 0 0 0
{P5, P4} 0 0 0 0 0

5
Table 3: Prediction results of second-order Markov Model.
Last web page Original web page Predicted web page Correct web page
{P3, P1} P4 P5 X
{P1, P5} P4 P4 √
{P4} P5 Cannot be predicted No Result
Table 3 shows the prediction results of second-order Markov Model. The prediction accuracy is
33 %. Test session P4 cannot be predicted with the help of second-order markov model. The
coverage of the model becomes 50% because it cannot predict the single state.
B) Generation of Dynamic Nested Markov mode
In Dynamic Nested Markov Model the higher-order Markov Model is nested inside the lower-
order Markov Model [13, 14]. DNMM uses the link list structure for storing the information of
web page. DNMM is same as Markov Model with some changes so that the efficiency of model
can be enhanced. This model is dynamic in nature means the addition and deletion of state can be
done easily. This model uses the node structure to store the web page. All the information of a
particular web page is stored in a node of that web page. In this model, only one node per web
page is created. Node is a dynamic data structure rather than just name of the web page. Each
node contains name of web page and an in-link-list. The inlink list is a link list in which each
node contain name of a previous web page from which the current web page is traversed, count
that shows number of times current web page is traversed from previous web page and an outlink
list that keep track of all the corresponding to that previous web page. Outlink list is a linked list
whose each node contains name of next web page and its count. Now this data structure keeps
track of all the previous web pages and all the next web pages corresponding to each previous
web page of the current node. In third order model every node contain data upto third order and in
fourth order each node contain data up to fourth order, but number of nodes are always constant.
In this way, we can model user navigation sessions in highly structured and efficient way.
In DNMM each web page is represented by unique node. All the information regarding a
particular web page is stored inside that node up to the n-order model. Figure 2 shows the node
structure of web page Wx for second-order DNMM. W1, W2…. are the second-order inlinks to
the web page Wx and from Wx the corresponding second-order outlinks are shown in figure 2.

6
Fig. 2: Node structure of Second-Order Dynamic Nested Markov Model.
Figure 3 shows the first-order DNMM corresponds to the table 1. S and E i.e. start and end is the
two artificial states incorporated in model. First-order DNMM construction starts with traversing
the first user navigation sessions of web log. If the traversed web page does not exists then the
node of that web page is created and corresponding f_inlink count and f_outlink list which has
next web page name and count will be created. Firstly, P2, P3, P1, P5 will be traversed because it is
the first sessions and there is no node existing previously. Node of P2 web page will be created,
f_inlink count set to 1 and f_outlink list which contains P3 as next web page name and count is 1.
Next P3 web page is created, f_inlink count set to 1, and f_outlink list will have P1 as next web
page name and count is set to 1. Now P1 web page will created, f_inlink count is set to 1 and
f_outlink has P5 as next web page and count is 1. This way the first-order DNMM is created. The
first-order Model DNMM is almost like the traditional first-order Markov Model.
Fig. 3: First-Order Dynamic Nested Markov Model corresponds to the training file of table 1.

7
Figure 4 shows second-order Dynamic Nested Markov Model corresponds to the table 1. In the
second-order DNMM history of previous web page is stored inside the node of first-order model.
The model construction starts with the first user navigation sessions i.e. P2, P3, P1 and P5. As it is
the first user navigation sessions of the training file and no node exists till now, so the P2 node
will be created. inlink of P2 node is created and named as S because P2 is starting web page and
S is considered to be its previous web page. Now the outlink list of inlink S will be updated as P3
web page and its count set to 1. Now P3 is traversed which is also not exists in the model so it is
created. P2 is created as inlink and count set to 1 because P2 is traversed one time from P3. Now
the outlink P1 for inlink P2 is created and its count set to 1. Similarly node P1 will be created.
When P5 node is created, its inlink P1 is created which has E as outlink. This way the first user
navigation sessions has been modelled. Second user navigation sessions is P2, P1, P3, P4, P5. Its
first web page is P2 which is already exist in the model. It has S as its inlink so count will be
incremented by 1 and will become 2. Outlink P1 will be checked in inlink of S which does not
exist, so it is created and count set to 1. Web page P1 is traversed and node P1 is present in model.
Node P1 has an inlink P3. Another inlink P2 is created and its outlink P3 is created. Similarly, web
page P3 is traversed and its node is updated. When web page P4 is traversed its node does not
exist, so it will be created and updated similarly. Remaining user navigation sessions are
modelled in same way.
Fig. 4: Second-order Dynamic Nested Markov Model corresponds to example of table 1.

8
Table 4: Prediction results of second-order DNMM.
Last web page Original web page Predicted web page Correct web page
{P3, P1} P4 P5 X
{P1, P5} P4 P4 √
{P4} P5 P5 √
Table 4 shows the Prediction results of second-order DNMM. There are three test sessions out of
which two were predicted right. The prediction accuracy is 66% because it can predict the single
state also. That is why its coverage is 100%.
4. Experiment Result:
Data Sets:
The experimental data set were obtained from three different data sources. First experimental
weblog data is collected from Cuboid Pvt. Ltd., Indore. Second weblog data is MSNBC, collected
from UCI repository and can be downloaded from
https://archive.ics.uci.edu/ml/datasets/MSNBC.com+Anonymous+Web+Data. Third experimental
weblog data were obtained from the authors Jose Borges and Mark Levene and downloaded from
http://www.cs. washington.edu/homes/map/adaptive/download.html. Two months web log data
are obtained from this website for experiments. The web site is http://machines.hyperreal.org. It
is given that web site receives approximately 10000 requests per day from around 1200 users.
Table 5 summarizes the characteristics of the web log data sets. The training data set and testing
data set characteristic are given below.
Table 5: Summary of Web Log Datasets (A) Training Data Set (B)Testing Data Set.
Training set Data Set Web pages Session Requests
Cuboid 29 3798 9566
MsnBc 92 3234 12378
HyperReal 36 2567 8821
(A)
Testing set Data Set Web pages Session Requests
Cuboid 29 3471 7894
MsnBc 92 2931 9087
HyperReal 36 2130 7845
(B)

9
Evaluation Parameters:
Various orders of Markov Model and Dynamic Nested Markov Model on three different datasets
have been evaluated. Evaluation parameters are Model Generation Time, Prediction Time,
Prediction Accuracy and Coverage.
(i)Model Generation Time
Generation time is defined as the time required by prediction model for modelling the training
file. Fig. 5 shows the model generation time of various order of Markov Model. Model generation
time is measured in millisecond. It is also observed that generation time depends on the number
of states generated by model. Second-order markov model generated more number of states as
compare to first order markov model so it take more time to generate.
Fig. 5: The time taken in millisecond for generation various order Model
(ii)Prediction Time
Fig. 6 shows prediction time taken by various order of model which is measured on the testing
file. The prediction time is depends on the other factor like the network traffic, server etc. But the
major time is of model which is observed in millisecond. The prediction time will be affected by
the number of web pages in the web log file also. The result shows that the prediction time will
increase with respect to number of web pages. The first-order Markov Model and DNMM take
same time to predict the next accessed web page. As move towards higher-order of Markov and
DNMM models, DNMM takes less prediction time as compared to Markov Model.

10
Fig. 6: The prediction time taken in millisecond for various order of model.
(iii)Prediction Accuracy
Prediction accuracy is very important parameter for the prediction model. It measures the
accuracy of the prediction model applied for testing file and calculated as:
Prediction Accuracy = (Number of correct prediction) / (Number of test sessions)
Where number of correct prediction are the number of test user navigation sessions which are
correctly predicted and number of test sessions are the total number of test sessions on which
prediction is performed. Correct predictions are find out by comparing original and predicted web
pages, those predicted web pages which are equal to original web pages are consider as correct
prediction.
As shown in figure 7, first-order Markov Model and DNMM gives same prediction accuracy. As
move towards higher-order of Markov and DNMM models, DNMM gives high prediction
accuracy as compared to Markov Model.
Fig. 7: The prediction accuracy for various order models.

11
(iv)Coverage
The coverage of the model is defined as the ratio of number of times model is able to predict to
number of requests in test set. In case of first-order Markov Model the coverage of state is 100%
and in second-order Markov Model the coverage of the state is 50% and as we move for the
higher order Markov Model the coverage will be less. For example in second-order Markov
Model where each state is a set of two web pages, if we want to predict that after web page
W1which will be the next web page accessed then model will fail to predict because it is not
having single state web page. In the DNMM in each order of the model, coverage will be 100%.
Conclusion
Various models have been analyzed in this work on the basis of Generation Time, Prediction Time
Prediction Accuracy and coverage. DNMM gives better prediction accuracy as compare to the
Markov Model. The coverage of the DNMM is better than the Markov Model.
References
[1] Federico Michele Facca, Pier Luca Lanzi, Mining interesting knowledge from weblogs: a survey,
Data & Knowledge Engineering, 53 (2005) 225–241, doi:10.1016/j.datak.2004.08.001
[2] S. Zhang and Z. Shi “Research and Application in Web Usage Mining of the Incremental Mining
Technique for Association Rule” IFIP International Federation for Information Processing, Springer
Boston, 2005.
[3] R. Popa, T. Levendovszky “Marcov Models for Web Access Prediction” 8th
International Symposium of Hungarian Researchers on Computational Intelligence and Informatics,
Nov 2007.
[4] V.Valli Mayil, “Web Navigation Path Pattern Prediction using First Order Markov Model and Depth
first Evaluation” International Journal of Computer Applications (0975 – 8887) Volume 45– No.16,
May 2012
[5] F. Khalil, J. Li and H. Wang “A Framework of Combining Markov Model with Association Rules for
Predicting Web Page Accesses”, Proc. Fifth Australasian Data Mining Conference (AusDM2006),
vol. 61, pp 177-184, 2006.
[6] J. Borges, “A Data Mining Model to Capture User Web Navigation. PhD Thesis”, University College
London, London University, 2000.
[7] J. Borges and M. Levene, “A Clustering-Based Approach for Modelling User Navigation with
Increased Accuracy”, Proc.Second Int’l Workshop Knowledge Discovery from Data Streams, pp. 77-
86, Oct. 2005.
[8] Borges, J. and Levene, M. 2005. Generating Dynamic Higher-Order Markov Models in Web Usage
Mining. Proc. Ninth European Conf. Principles and Practice of Knowledge Discovery in Databases
(PKDD). A. Jorge, L. Torgo, P. Brazdil, R. Camacho, and J. Gama, Eds. (Oct. 2005), 34-45.
[9] Mathias G´ery, Hatem Haddad, “Evaluation of Web Usage Mining Approaches for User’s Next
Request Prediction” November 2003, New Orleans, Louisiana, USA. Copyright 2003 ACM.
[10] Khalil, F., Li, J. and Wang, H. Web Path Recommendations Based on Page Ranking and Markov
Models. Proc. Seventh Ann. ACM Int’l Workshop Web Information and Data Management (WIDM
’05). (2005), 2-9. 2006.
[11] Siriporn Chimphlee, Naomie Salim, Mohd Salihin Bin Ngadiman and Witcha Chimphlee, Using
Association Rules and Markov Model for Predit Next Access on Web Usage Mining” 2006, Advances
in Systems, Computing Sciences and Software Engineering, Pages 371-376

12
[12] Nizar R. Mabroukeh, and C. I. Ezeife, “ Semantic-rich Markov Models for Web Prefetching, in the
proceedings of the 2009 IEEE International Conference on Data Mining (ICDM) Workshops
(Workshop on Semantic Aspects in Data Mining (SADM'09)), Miami Florida, December 6-9, 2009,
pp. 465–470.
[13] B. Nigam, S. Jain, "Generating a New Model for Predicting the Next Accessed Web Page in Web
Usage Mining", in Proceedings of the 3rd International Conference on Emerging Trends in
Engineering and Technology (ICETET ’10), Washington, DC, USA, pp.485-490, 2010.
[14] Bhawna Nigam and Suresh Jain “Analysis of Markov Model on Different Web Prefetching and
Caching Schemes” ICCIC, 28-29 Dec—2010 (2010 IEEE International Conference on Computational
Intelligence and Computing Research (ICCIC)), Tamilnadu College of Engineering Coimbatore,
ISBN: 978-1-4244-5965-0. Digital Object Identifier: 10.1109/ICCIC.2010.5705732
[15] http://archive.ics.uci.edu/ml/datasets/MSNBC.com+Anonymous+Web+Data
[16] J. Borges and M. Levene, Data Mining of User Navigation Patterns. Web Usage Analysis and User
Profiling, B. Masand and M. Spiliopoulou, Eds. LNAI 1836, 2000, 92-111. Springer.
[17] M.T. Hassan, K.N. Junejo, and A. Karim, “Learning and predicting key Web navigation patterns using
Bayesian models,” in Proceedings of Int. Conf. Comput. Sci. Appl. II, Seoul, Korea, pp. 877–887,
2009.
18] Mamoun A. Awad and Issa Khalil, “Prediction of User’s Web-Browsing Behavior: Application of
Markov Model,” IEEE Trans. Syst., Man, Cybern. A., Syst., Humans, Volume 42, No. 4, pp., Aug.
2012.
[19] Poornalatha G, Prakash S Raghavendra, “Web Page Prediction by Clustering and Integrate Distance
Measures” IEEE/ ACM Trans. Syst., Man,Cybern. A, Syst., Humans, Volume 44, No. 2, pp., Sep.
2012.

Evaluation of models for predicting user’s next request in web usage mining

More Related Content

What's hot (17)

Viewers also liked (19)

Similar to Evaluation of models for predicting user’s next request in web usage mining (20)

Recently uploaded (20)

Evaluation of models for predicting user’s next request in web usage mining