International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)

Web Site: www.ijettcs.org Email: [email protected], [email protected]

Volume 3, Issue 2, March April 2014 ISSN 2278-6856


Volume 3, Issue 2 March April 2014 Page 145


Abstract: Classification technique is important concept of
web usage mining useful for researchers in web prediction
where users browsing behavior is predicted. The knowledge
of navigated web pages history recorded from user browsing
sequence at web log server which is used to predict future set
of pages likely to be visited by user. Lots of research has been
carried out on prediction of next page Request for various
kinds of applications including wireless application, caching
system, search engine, personalization and recommended
system. The main objective of this paper is to focus on the
survey and analysis of users future page request prediction
for recommended system. The detailed review is presented on
various kind of prediction models used for Future page
Request Prediction. The study and contrast of prediction
models will give a new way for researchers to go deeper to
find a pearl from shell in the sea.

Keywords: Prediction models, Web prediction,
predicting factors.

1. INTRODUCTION
Classification technique is important concept of web
usage mining helpful for researchers in web prediction
where users browsing behavior is predicted. Users
browse the web pages, the history of browsed pages
recorded in logs at server side. The recorded information
in log is helpful to predict users web browsing behavior.
The knowledge of navigated web pages history is used to
predict futures set of pages likely to be visited by user.
This kind of prediction can be applied in various
applications such as recommended system, wireless,
caching, search engine, etc. The objective of this paper is
to reviewing different prediction models which focuses on
prediction of future set of pages user may visit from
recorded web log history.

The ideas and solutions introduced in this survey may
motivate the recommended system developer to turn
research into reality.

The organization of this paper is as follows: In section II,
presented the motivation behind this survey. In section
III, we introduce the literature survey of various models
proposed by researchers for recommendation. Section IV
is the summary of the survey and last is acknowledgement
and references.

2. Motivation
Internet is source of informative world where users
considered it as a first home to motivate and take the
guidance of intelligence for their desire purpose. In
todays era of Internet web-browsing is common in
human being and often users access the web worldwide.
Web browsing is enlightening and descriptive tool to gain
the impression of contents for the users to achieve their
desire goals and objectives. Users surfing activities of the
web are in large amount of data recorded in log files
stored at server side. By identifying browsing patterns of
users from log files and predicting future request pages
for recommended system is the main motivation.

3. LITERATURE SURVEY
The main focus of literature survey is to study and
contrast the prediction models to predict the users future
web page requests. Prediction models are used addressing
web prediction problem. The main aim is to study
different prediction models to reduce users access times
and improving personalization while browsing the
services of web. In addition, to reduce network traffic
problems by avoiding pages visiting unintentionally and
unnecessarily by users. Various prediction model like
Markov model, artificial neural networks (ANN), k-
nearest neighbor (kNN), support vector machine (SVM),
fuzzy inference, Bayesian model are proposed by
researchers to predict user future request of page.

Prediction models can be classified into two categories
named as point-based and path based prediction models.
When users previous and historic path data are predicted
then it is referred as path based prediction. Point-based
prediction is based on users current actions.

Joachims et al. [1] proposed a tour guide called web-
watcher which is a path based recommender model based
on kNN and reinforcement learning. Web watcher is a
tour guide agent helps the users from their browsing
interest by providing the directions to highlight selected
hyperlinks from original page. Highlighting is done by
inserting eyeball icon around the hyperlinks. Web
watcher is an interfacing agent between the user and the
World Wide Web as shown in figure 1.
Review on Exploring Users Surfing Behavior
for Recommended Based System

Mr. Ajeetkumar S. Patel
1
, Prof. Anagha P. Khedkar
2


1
Department of Computer Engineering,
Matoshri College of Engineering and Research Center,
Nashik, India

2
Department of Information Technology,
Matoshri College of Engineering and Research Center,
Nashik, India
International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)
Web Site: www.ijettcs.org Email: [email protected], [email protected]
Volume 3, Issue 2, March April 2014 ISSN 2278-6856


Volume 3, Issue 2 March April 2014 Page 146



Figure 1 Web watcher: Interface agent between user and
World Wide Web [1]

The function of web watcher is to suggest appropriate link
given an interest and webpage. To fulfill this need
requires knowledge of following target function.

LinkQuality: Page X Interest X Link [0, 1] [1]

This gives the value of LinkQuality interpreted as the
probability that a user will select the Link give the current
page & Interest.
Web watcher used the k nearest neighbor model for
knowledge discovery by wishing the value of LinkQuality
for each hyperlink estimated to be the average similarity
of the k (considered as 5) highest ranked keyword sets
associated with the hyperlink. The hyperlink value for
LinkQuality is above threshold then it is suggest as user
interest.
Reinforcement learning is second approach used in web
watcher agent allows learning control strategies that
select optimal actions in certain settings.
The main objective is to find paths through the web to
maximize the amount of relevant information
encountered [1]. Web watcher experimentation is
considered in all cases of recommendation give only 48%
throughput because of great diversity of individual users
interests. The prediction accuracy will be considered as
future research and scope to know more deeply in users
interest.
Nasraoui et al. [2] proposed a web recommended system
by applying clustering to group profiles using hierarchical
unsupervised niche clustering (HUNC). HUNC algorithm
is used to cluster the sessions extracted from server logs
into multi-resolution user session profile and also identify
the noisy sessions and profiles. Moreover, clustering
process hunt the context-sensitive associations between
different URL addresses.
The main advantage is no additional cost required as well
as it can be generalized to other transactional data.
HUNC gives faster performance and in addition by
applying genetic algorithms parallelization efforts make it
even faster.
Mobasher et al. [3] proposed association rule mining
(ARM) technique and recommended system framework to
predict users future page that likely to be visited. To
match an active user sessions with frequent item sets can
be presented by the frequent item set graphs to predict
next pages. Framework is categorized in three phases:
data preparation, pattern discovery and recommendation
engine.

Table 1: Web transactions involving page view [3]
T1: {p1 p2 p3 p4 p5}
T2: {p1 p2 p5 p3 p4}
T3: {p1 p2 p5p p3}
T4: {p2 p5 p2 p1 p3}
T5: {p4 p1 p2 p5 p3}
Table 2: Frequent itemsets generated by the Apriori
algorithm [3]
Size 1 Size 2 Size 3 Size 4
{p1}(5)
{p2}(6)
{p3}(4)
{p5}(5)
{p1, p2} (5)
{p1, p3} (4)
{p1, p5} (5)
{p2, p3} (4)
{p2, p5} (5)
{p3, p5} (4)
{p1, p2, p3}(4)
{p1, p2, p5}(5)
{p1, p3, p5}(4)
{p2, p3, p5}(4)

{p1, p2,
p3, p5} (4)

Data cleansing, user identification, session identification,
page view identification and transaction identification is
performed on data preparation phase. This preprocessing
tasks ultimately result in a set of n page views, P ={p1,
p2, p3, .., pn}, and set of m user transactions T ={t1,
t2, t3, .. , tn} where each ti T is a subset of P [3]. In
pattern discovery phase, an association rule captures
relationships among page views based navigational
patterns of users.
Apriori algorithm is used to find group of items occurring
frequently together in many transactions referred as
frequent item sets. The last phase takes a collection of
frequent item sets as input and generates recommendation
set. This set is for user by matching the current users
action against training patterns. Recommendation engine
is online process framework includes data structure for
storing frequent item sets combined with recommendation
algorithm to directly produce real-time recommendations
from item sets without first generating association rule.
For understanding the concept following example is
illustrated. Table 1 shows transaction sets. Apriori
algorithm generates itemsets using transactions given in
table 1 by considering frequency threshold of 4 shown in
table 2. By referring table 2 following figure 2 shows
frequent itemsets graph based on frequent itemsets.

Figure 2 Frequent itemset for an example [3]

Association rule mining is very much suffering from
scalability and efficiency in web prediction give path to
research on it.
International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)
Web Site: www.ijettcs.org Email: [email protected], [email protected]
Volume 3, Issue 2, March April 2014 ISSN 2278-6856


Volume 3, Issue 2 March April 2014 Page 147


Anderson et al. [4] proposed adaptive learning
MINIPATH algorithm using dynamic links to provide
shorter path to save substantial surfing/navigational
efforts. The proposed system specifically designed for
wireless gadgets PDAs, cell phones, tablets and pagers
used by users to browse the web. In offline mode
MINIPATH learns the history of users browsing
behavior and used at runtime to predict the users
ultimate destination. Researcher also has taken the
interest in the evaluation of some prediction models of
web usage mining with MINIPATH such as nave bayes
mixture model and Markov mixture model.
The resultant output of the mixture of Markov model save
more than 40% navigational efforts of users. Proposed
system is able to reduce the visitors links of browsing but
up to specific mark. Beyond that MINIPATH is not able
to maintain the consistency. The reason behind it is
unable to handle larger sites having more pages with lots
of link in between them. The cost of predicting shortcut
link with probability of erroneous to follow a shortcut is
major issue in MINIPATH. The accuracy of mixture
Markov model has somewhat needs the improvement.
Perkowitz et al. [5] give the challenge to AI community to
create adaptive websites while user accesses the data and
then automatic improvement is desired on organization
and presentation of the sites. Proposed system suggested
two ways for adaptive websites: customization and
optimization. Modifying pages of web at runtime to suit
the needs of individual users is referred as customization.
Changing the site itself to make navigation easier for all
is referred as optimization. To do so there is need of
users access patterns which will be avail from server logs
but for both operations available information is
inadequate. So there is a need of meta-information to
support adaptivity.
Su et al. [6] proposed the N-gram prediction model to
utilize path profiles of users from very large server log
datasets to predict users future set of page requests. Two
models are introduced: path-based prediction model built
on web-server log and N-gram prediction model
constructed out of log data to predict users clicks in real
time. To elaborate the proposed model refers following
example. Consider a log file consisting following request
paths given in table 3.

Table 3: Example of log file consisting a request path [6]
p1, p2, p3, p4
p1, p2, p3, p5
p1, p2, p3, p5
p2, p3, p4, p6
p2, p3, p4, p6
p2, p3, p4, p5
Table 4: 3-gram prediction model H3 () for
an Example [6]
N-gram Prediction
p1, p2, p3 p5
p2, p3, p4 p6
Table 5: 2-gram prediction model H2 () for
an Example [6]
N-gram Prediction
p1, p2 p3
p2, p3 p4
p3, p4 p6

First approach the path-based model which can be used to
build an N-gram prediction model based on frequency
occurrence. To build 3-gram model on paths (table 3)
gives two 3-grams (p1, p2, p3 and p2, p3, p4) to build
prediction model. Algorithm return given in hash table 4
called H3 ().
2-gram model provide contents while building it: p1, p2;
p2, p3 and p3, p4. 2-gram prediction model H2 () can be
return given in table 5.
Second approach prediction algorithm (N-gram+) is used
to build prediction model. 3-gram H3 () and 2-gram H2 ()
models are the example of it. By an observation of above
example if current clicking sequence consists of only 1
click p4p2p3. Prediction algorithm first checks H3 ()
for the existence of index p4p2p3. If index doesnt exist
then it will check 2-gram model H2 () for index p2p3.
p2p3 is exist thus the predicted next click is p4 as per
H2 ().
These methods give the performance on increasing in
precision and decreasing in applicability. This work
shows better prediction accuracy while maintaining
reasonable applicability using N-gram models where n is
greater than 2. N-gram+ algorithm only applicable to
significant portion of the logs may degrade the system
performance on prediction of users access patterns as
well as on prediction time.
Levene et al. [7] presented Markov chain model for
analysis of users web browsing behavior by computing
information contained in navigational trail. Main result is
carried out in twofold: First is to obtain a large deviation
result for the trail of web trough Markov chain needed to
compute its entropy in terms of the mean waiting and
recurrence time between states. Second is to present an
online algorithm to compute entropy of Markov chain to
converge to the true entropy. In addition an error analysis
is provided for said algorithm in terms of the relative
entropy between empirical and actual transition
probabilities.

Table 6: Algorithm to estimate entropy of
Markov chain [7]
Algorithm 1 (ENTROPY ({m
i
}, {m
i, j
})) [1]
1. begin
2. H, {m
i
}, {m
i, j
}, k :=0;
3. s
1
:=random({P
i
});
4. while not CONVERGED do
5. k :=k +1;
6. s
k+1
:=random(s
k
; {P
kj
});
7. H :=H +(m
k,(k+1)
+1) log((m
k,(k+1)
+1)/( m
k
+1))
m
k,(k+1)
log(m
k,(k+1)
/m
k
);
International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)
Web Site: www.ijettcs.org Email: [email protected], [email protected]
Volume 3, Issue 2, March April 2014 ISSN 2278-6856


Volume 3, Issue 2 March April 2014 Page 148


8. for all j (k + 1) do
9. H :=H +m
k,j
log(m
k
=(m
k
+1));
10. end for
11. m
k
:=m
k
+1;
12. m
k, (k+1)
:=m
k, (k+1)
+1;
13. end while
14. return H/k;
15. end.

Table 6 shows algorithm ENTROPY ({m
i
}, {m
i, j
})
returns an estimate of entropy of the Markov model M by
overloading index k of state s
k
so that it refers both to the
kth state in the trail induced by the random walk and also
to the kth state in N.
{m
i
} and {m
i, j
} indicates number of visits to s
i
and the
number of transitions from s
i
to s
j
so far, where i, j {1,
.., n}. CONVERGED is Boolean variable referred as
initially false [7]. Function random ({P
i
}) is used to
return an initial state distribution according to the initial
probabilities {P
i
}. Function random (s
i
, {P
ij
}) returns a
state adjacent to s
i
distributed according to the transition
probabilities {P
ij
} where i, j {1, .., n}.
Pitkow & pirolli [8] proposed the technique called longest
repeating subsequences (LRS) which reduces the
modeling complexity of surfing patterns of users. To
understand LRS concept consider following example.
Suppose websites contains the pages p1, p2, p3 and p4. In
that p1 contains hyperlinks to p2 and p2 contains
hyperlinks to both p3 and p4 refer following figure 3.
In case 1 if users repeatedly surf from p1 to p2, but only
one user clicks through to p3 and one user clicks through
p4.

Figure 3 Example illustrating the formation of Longest
Repeating Subsequences [8]

So from this we conclude that the longest repeating
subsequence is p1p2. In case 2 if more than one user
clicks through p2 to p4, then both p1p2 and p1p2p4 are
longest repeating subsequences. If both p1p2p3 and
p1p2p4 are the Longest Repeating Subsequences when
both occur more once and are the longest subsequences is
case 3. In case 4 if user click through p1p2p4 then p1p2
is not a longest repeating subsequence.
Summarizing longest repeating subsequences gives
several interested properties. First, low probability
transition removal gives complexity reduction. Second,
LRS having bias towards specificity will give good impact
on hit rate. Experimentation of LRS gives clues that need
improvement in adaptation and real-time model.
Proposed work give future steps based on loss of pattern
matching as well as prediction accuracy on large dataset
input.
Hassan et al. [9] present Bayesian model for web
navigation patterns focusing on long sessions (short and
long), kind of pages, page views range, and page
categories rank problem.
Web navigation patterns are carried out by learning and
prediction for user, time spend for surfing and user-time
slot combinations employing bayes rule and Markov
chain models to gain satisfaction of accuracy and
simplicity. Bayesian model is adapted for following
reasons. 1) Simple and intuitive 2) adaptability to concept
drift 3) efficiency and accuracy. Table 7 shows four
pattern problems and prediction performance [9].

Table 7: Classifier/model used for patterns [9]
Pattern
No.
Pattern Name
Classifier/
technique
1
Short and long visit
sessions
Nave bayes
classifier
2
Page categories visited
in first N positions
Markov chain
3
Range of number of
page views per page
category visited
Bayes
classifier
4
Rank of page categories
in visit sessions
Nave bayes
classifier

Hence, we can say that proposed model is generative in
nature and for pattern1 evaluation doesnt improve
prediction performance. Complexity of for all patterns is
O (D) in time, where D is total number of visit sessions in
the data. The space complexity is sum of the products of
variables cardinalities as defined by the number of
probability required. Proposed model for learning and
predicting sessions has same prediction accuracy but
faster in process than support vector machine.
Awad et al. [10] proposed the fusion of several prediction
models, namely, Markov model and support vector
machines with domain knowledge exploitation to improve
prediction accuracy. Proposed system is divided in to two
sources of prediction support vector machine (SVM) and
Markov model. In SVM two ideas are considered for
reality. First, extraction of users browsing sessions
features and trained it with SVM. Second is by
integrating domain knowledge in prediction to improve
prediction accuracy.
The second body of prediction is probabilistic model
(Markov model) utilizing the best of SVM and Markov
model which surpass prediction accuracy all other
techniques. But problem occurs on proposed models is
prediction and training overhead. ARM and SVM model
has some limitations as scaling on large datasets.
Whereas SVM and ANN having limitations of handling
multiclass problem efficiently due to larger number of
labels/classes involved in web prediction technique.
Awad et al. [11] synthesis the classification techniques
namely, Markov model, ANN and All- Kth Markov
model to resolve prediction using Dumpsters rule which
indeed boosting the accuracy on large number of classes
International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)
Web Site: www.ijettcs.org Email: [email protected], [email protected]
Volume 3, Issue 2, March April 2014 ISSN 2278-6856


Volume 3, Issue 2 March April 2014 Page 149


of ANN and by engaging reduction technique with
domain knowledge to enhance result by reducing number
of classifiers to improve prediction accuracy and
prediction time of ANNs. The fusion of prediction models
shows outperformed results on prediction accuracy but
compromises the scalability in number of paths and
fusion may compromise on prediction time.
To overcome the limitations of [10] [11] Awad et al. [12]
proposed modified Markov model and novel two-tier
prediction framework to gain better result in prediction
accuracy (by reducing the complexity of original Markov
model) and prediction time (without compromising
prediction accuracy) respectively. Experimentation was
done on the three standard benchmark datasets, namely,
the NASA datasets, the University of Saskatchewans
(UOFS) dataset and the United Arab Emirates University
(UAEU) datasets. [13].
The proposed model operations needs background study
and analysis of N-gram representation of sessions,
Markov model, All- Kth Markov model, ARM and utilize
them for same. N-gram representation is used in web
prediction to represent training session by depicting
sequences of page clicks by users navigating a website [6]
[12]. Sliding window method is used at the time of
implementation and training the session. Markov model
is used to implement the web prediction concept by
predicting next page likely to be visited based on history
of previous pages visited for random order of k to achieve
efficiency and performance on model building and
prediction time. In fourth-order Markov model, the
prediction of the future page is computed based only on
the four web pages previously visited [2] [3]. In All Kth
Markov model all orders of Markov model are used
collectively in prediction to increase accuracy [2] [3]. One
of the data mining technique is used called ARM for
solution of web prediction by providing support on the
accuracy and maintain relationship of varying the number
of items.

Figure 4 Different phases in two-tier prediction
framework training process [12]

Proposed modified Markov model is introduced to
achieve better memory utilization and faster prediction by
reducing number of paths in the model. The main
objective behind it is users browsing order on the web.
Discarding the repeated page sessions of browsing order
is being carried out on observation based. Second
approach is two-tier prediction framework that creates an
example classifier (EC) based on the training examples
and generated classifiers to improve prediction time. The
noticeable point is no compromise on prediction accuracy
[12]. Following figure 4 shows the two-tier framework
training process.
In Proposed model, first N orders of Markov models are
generated, namely, first, second, .., Nth order Markov
model by applying sliding window on training set T, then
mapping each training example in T to one or more
orders of Markov models. In figure 4 t3 training example
is mapped to two classifier IDs namely C1 and C2 as first
order and second order Markov model respectively.
Checking highest probability of classifiers examples
mapped to only one classifier in pruning/filtering process
and at last EC is generated based on training the filtered
dataset using one of the prediction techniques. SVM is
used to generate EC because there is a need of any
learning techniques to generate EC may include ANNs or
decision trees.
Extracting statistical features researchers may improve
prediction accuracy by giving improvement over proposed
system. Furthermore using ANN on error back
propagation instead of SVM to generate example
classifiers may give better outcome to improve
performance of model.

4. SUMMARY
In this paper, survey is presented on several kinds of
prediction models for web prediction problem especially
focusing on improving the effectiveness of web prediction
for recommendation system on prediction accuracy,
prediction time, performance, efficiency, simplicity,
memory utilization, applicability and scalability by
session extraction of web logs, user understanding, model
understanding, document understanding with ranking
and monitoring of evaluation and feedbacks. Many
advanced prediction model were developed as well as
mixture of them also utilized. Those models were either
applied to large amount of log datasets or part of datasets
for web prediction to enhance the powerfulness of
existing work. There is still challenging and interesting
problems present in existing systems. Challenges may be
categorized two sections namely preprocessing and
prediction. Those challenges are highlighted by following
points mentioned below.
Domain knowledge hunt.
Handling large amount of data that can be fit in
computer memory to improve memory utilization.
Session identification to ease the working of
proposed methods.
Improvement over long training dataset timing.
Improvement over prediction time.
Selection of prediction model.
Hybrid approach involvement for improvement on
prediction time.
International Journal of EmergingTrends & Technology in Computer Science(IJETTCS)
Web Site: www.ijettcs.org Email: [email protected], [email protected]
Volume 3, Issue 2, March April 2014 ISSN 2278-6856


Volume 3, Issue 2 March April 2014 Page 150


Prediction accuracy.
Modeling complexity.
Relativity between prediction accuracy, prediction
time and modeling complexity.

Acknowledgement
Sincerely thank the all anonymous researchers for
providing us such helpful opinion, findings, conclusions
and recommendations. Also thanks to guide Prof. Anagha
P. Khedkar, HOD Dr. Varsha H. Patil, Principal Dr. G.
K. Kharate, and colleagues for their support and
guidance.

References
[1] T. Joachims, D. Freitag, and T. Mitchell,
WebWatcher: A tour guide for the World Wide
Web, in Proc. IJCAI, 1997, pp. 770777.
[2] O. Nasraoui and R. Krishnapuram, One step
evolutionary mining of context sensitive associations
and Web navigation patterns, in Proc. SIAM Int.
Conf. Data Mining, Arlington, VA, Apr. 2002, pp.
531547.
[3] B. Mobasher, H. Dai, T. Luo, and M. Nakagawa,
Effective personalization based on association rule
discovery from Web usage data, in Proc. ACM
Workshop WIDM, Atlanta, GA, Nov. 2001.
[4] C. R. Anderson, P. Domingos, and D. S. Weld,
Adaptive Web navigation for wireless devices,
Proc. IJCAI Workshop, Seattle, WA, 2001.
[5] M. Perkowitz and O. Etzioni, Adaptive Web sites:
An AI challenge, in proc. IJCAI Workshop, Nagoya,
Japan, 1997.
[6] Z. Su, Q. Yang, Y. Lu, and H. Zhang, WhatNext: A
prediction system for Web requests using n-gram
sequence models, in Proc. 1st Int. Conf. Web Inf.
Syst. Eng. Conf., Hong Kong, Jun. 2000, pp. 200
207.
[7] M. Levene and G. Loizou, Computing the entropy of
user navigation in theWeb, Int. J. Inf. Technol.
Decision Making, vol. 2, no. 3, pp. 459476, 2003.
[8] J. Pitkow and P. Pirolli, Mining longest repeating
subsequences to predict World Wide Web surfing,
in Proc. 2nd USITS, Boulder, CO, Oct. 1999.
[9] M. T. Hassan, K. N. Junejo, and A. Karim,
Learning and predicting key Web navigation
patterns using Bayesian models, in Proc. Int. Conf.
Comput. Sci. Appl. II, Seoul, Korea, 2009, pp. 877
887.
[10] M. Awad, L. Khan, and B. Thuraisingham,
Predicting WWW surfing using multiple evidence
combination, VLDB J., vol. 17, no. 3, pp. 401 417,
May 2008.
[11] M. Awad and L. Khan, Web navigation prediction
using multiple evidence combination and domain
knowledge, IEEE Trans. Syst., Man, Cybern. A,
Syst., Humans, vol. 37, no. 6, pp. 10541062, Nov.
2007.
[12] M. Awad and I. Khalil, Prediction of Users Web-
Browsing Behavior: Application of Markov Model,
IEEE Trans. Syst., Man, Cybern. A, Syst., Humans,
vol. 42, no. 4, pp. 11311142, Aug. 2012.
[13] Internet Traffic Archieve. [Online]. Available:
http://ita.ee.lbl.gov/html/traces.html