DETECTING PARALLEL BROWSING TO IMPROVE WEB

PREDICTIVE MODELING

Geoffray Bonnin, Armelle Brun and Anne Boyer

LORIA – KIWI Team, B.P. 239, 54506 Vandœuvre-L`es-Nancy Cedex, France

Keywords:

Parallel browsing, Web recommendation, Predictive modeling.

Abstract:

Present-day web browsers possess several features that facilitate browsing tasks. Among these features, one of

the most useful is the possibility of using tabs. Nowadays, it is very common for web users to use several tabs

and to switch from one to another while navigating. Taking into account parallel browsing is thus becoming

very important in the frame of web usage mining. Although many studies about web users’ navigational

behavior have been conducted, few of these studies deal with parallel browsing. This paper is dedicated

to such a study. Taking into account parallel browsing involves to have some information about when tab

switches are performed in user sessions. However, navigation logs usually do not contain such informations

and parallel sessions appear in a mixed fashion. Therefore, we propose to get this information in an implicit

way. We thus propose the TABAKO model, which is able to detect tab switches in raw navigation logs and to

beneﬁt from such a knowledge in order to improve the quality of web recommendations.

1 INTRODUCTION

Predicting web users’ future paths is one of the most

important tasks in Web Usage Mining (WUM). WUM

can be deﬁned as “the process of applying data min-

ing techniques to the discovery of usage patterns from

web data” (Srivastava et al., 2000). Web predictive

modeling is useful for many purposes such as web

page research (Tan and Kumar, 2002), latency re-

duction (Schechter et al., 1998), arrangement of the

links in a website (Chi et al., 1998), web recommen-

dation (Nakagawa and Mobasher, 2003), etc. The

most popular techniques used in this domain are as-

sociation rules (Agrawal et al., 1993), sequential pat-

terns (Agrawal and Srikant, 1995) and Markov mod-

els (Pitkow and Pirolli, 1999). These techniques mod-

elize users’ sessions by exploiting a training corpus,

processed from navigation logs. These logs usually

consist in sequences of resources along with times-

tamps and users’ IP addresses, or ids. The usual

challenge is to provide a model that is a tradeoff be-

tween predictive accuracy, coverage, space and time

complexity (Pitkow and Pirolli, 1999; Deshpande and

Karypis, 2004).

Although widely studied, this domain is con-

stantly evolving, and is perpetually challenging. In-

deed, the technology used to surf on the web be-

comes more ergonomic and sophisticated everyday,

which involves changes in the usual users’ behaviors.

In particular, present-day browsers provide interfaces

called tabs that allow several pages to be contained

in the same window and to switch from one to an-

other, which is commonly referred to as tabbing. Al-

though this phenomenon is very common in current

web users’ behavior, very few approaches have dealt

with it. In several works, it is assumed that user ses-

sions include several tasks that can be characterized

(Jin et al., 2005; G¨und¨uz and

Ozsu, 2003). In one

session, a user can perform several tasks by tabbing,

which is referred to as parallel browsing. Tabbing ac-

tions are not stored in the navigation logs, only the

time-ordered actions are stored, and when parallel

browsing is performed the resulting session consists

in a mixed sequence of resources (mixed tasks). On

the client side, tabbing can be categorized into two

categories: inter-site and intra-site tabbing. On the

server side, only intra-site tabbing can occur. Regard-

ing inter-site tabbing, it is easy to detect which pages

correspond to which task, as one can simply consider

the url associated with the resources. In the case of

intra-site navigation, it is more difﬁcult to detect tab-

bing. For instance, the typical way to perform such an

activity is to click on a link using the middle button of

the mouse, or to hold down the CTRL key while click-

ing on a link. As neither this action nor tab switch-

ing are indicated in the navigation logs, determining

504

Bonnin G., Brun A. and Boyer A..

DETECTING PARALLEL BROWSING TO IMPROVE WEB PREDICTIVE MODELING.

DOI: 10.5220/0003115905040509

In Proceedings of the International Conference on Knowledge Discovery and Information Retrieval (KDIR-2010), pages 504-509

ISBN: 978-989-8425-28-7

 2010 SCITEPRESS (Science and Technology Publications, Lda.)

which resource within a session corresponds to which

task is a difﬁcult problem. State-of-the-art predictive

models do not take into account tabbing, they usu-

ally simply assume that one session is made up of one

single task. We assume that the prediction ability of

these models should be increased if the tasks were ex-

tracted from the sessions and considered for predic-

tion. In this paper we aim at discriminating between

the different tasks mixed within a session. We pro-

pose to tackle this issue by using sequence alignment

algorithms. We ﬁrst propose an algorithm to extract

tasks from sessions. We then propose a model that ex-

ploits the extracted tasks in order to perform resource

recommendation.

The rest of this paper is organized as follows. We

are ﬁrst interested in works related to parallel brows-

ing and task-level modeling. We then present our

proposition to extract linear sessions from raw logs

in which sessions are mixed. Then, we introduce the

model we propose to perform recommendation. Con-

clusion and perspectives are put forward in the last

section.

2 RELATED WORK

Web predictive modeling has been widely studied in

the past decade. In this section, we ﬁrst present a brief

overview of the most usual models. We then turn to

the works related to parallel browsing and multi-task

modeling.

2.1 Web Predictive Modeling Overview

Association Rules and Sequential Patterns. One

way of exploiting past actions to predict future ones,

is the use of association rules. Association rules

have been initially used for mining supermarket bas-

ket (Agrawal et al., 1993) to extract information about

purchased items dependencies. An association rule is

an expression of the form X → Y, where X and Y are

sets of items. X is called the antecedent and Y the

consequent. An association rule means that, in one

transaction, when users have purchased all resources

in X then there is a high probability that they will pur-

chase Y. Using association rules in the frame of web

usage modeling thus enables to take into account non-

ordered sets of resources in the history. Sequential

patterns (Agrawal and Srikant, 1995; Lu et al., 2005)

are the sequential form of association rules, and are

thus more constrained than association rules.

In both approaches, statistical models are built us-

ing a training corpus by counting the sets of resources

or sequences of resources. Then, during the predic-

tion step, all possible antecedents in current user’s

navigation history are compared to the antecedents in

the model. If some antecedents match, then the corre-

sponding consequents are recommended.

Sequential patterns can be contiguous (Jianyong

et al., 2007) or non contiguous (Ayres et al., 2002).

Looking for non contiguous sequences of unlimited

sizes induces a huge amount of combinations. So the

step of pattern discovery has to limit the size of the

patterns to discover. A sliding window with a ﬁxed

size is usually used during the pattern discovery step

as well as during the recommendation step. However,

the time complexity induced is still high. As they in-

duce less combinations, the time complexity of con-

tiguous sequential patterns is lower.

Markov Models. In the frame of web naviga-

tion Markov chains are used to predict the next re-

source according to the present state (the k previously

browsed resources), which is referred to as Markov

models of order k or k

-order Markov models.

Markov models are built in the same way associa-

tion rules and sequential patterns are, i.e. by browsing

a training corpusand counting sequences of size k+1.

The prediction step is similar to the one of sequen-

tial patterns too: the previous actions are compared

to the states in the model, and if some state matches,

then the most probable corresponding resource is pre-

dicted.

The performance of Markov models can be eval-

uated in terms of coverage, accuracy and state per-

plexity. Coverage is the percentage of cases where a

state of the model matches the current history. When

a large enough training data is available, higher or-

der Markov models provide a higher accuracy; how-

ever moving to higher ordersusually involves a higher

state complexity and a lower coverage, as no large

enough training data can be found in order to set val-

ues to each possible states. Thus, the use of Markov

models involves a tradeoff between accuracy, state

complexity and coverage (Pitkow and Pirolli, 1999).

One way to provide both high accuracy and high cov-

erage is to use several Markov models having various

orders. For example, one can try to provide a rec-

ommendation using a Markov model of order 3 (high

accuracy), and if no matching can be performed, try a

Markov model of order 2 (lower accuracy but higher

coverage), and so on, until a recommendation can be

provided. In the worst case, a Markov model of order

0 is used, which corresponds to the overall probability

of one single resource, without considering previous

resources. Using such a scheme, a full coverage can

be reached, while providing a good accuracy in the

recommendations. This scheme is called the all-k

DETECTING PARALLEL BROWSING TO IMPROVE WEB PREDICTIVE MODELING

505

order Markov model, and is one of the best perform-

ing predictive models of the state-of-the-art (Desh-

pande and Karypis, 2004). Notice that under the same

pruning conditions, it is similar to sequential patterns

in which patterns are contiguous.

One of the major drawback of the all-k

-order

Markov model is its low robustness to navigation mis-

takes and parallel browsing. Indeed, it takes into

account only strictly contiguous sequences, and if a

given user makes parallel navigations or goes to an

unwanted resource, the model will not be able to cap-

ture the real path and will reduce the size of the his-

tory considered. Moreover, when the model does not

match the complete history, the farthest resources are

discarded, and the closest resources are always con-

sidered while some of them may be navigation mis-

takes and should be discarded.

2.2 Parallel Browsing and Multi-task

Modeling

To the best of our knowledge, parallel browsing has

not yet been directly studied in the frame of web pre-

dictive modeling. However some works have dealt

with related interests.

In (Weinreich et al., 2006), a 195 days study of

user browsing behavior is presented. The authors

compare their study to two other user browsing be-

havior studies, that date back to 1995 and 1997 (no

other similar study has been performed during the

nine previous years). The results show important

changes in users’ behavior within these nine years,

such as a decreased use of the back button from 30%

in the mid-nineties to less than 15% currently. The use

of multiple browsing windows

has increased from

less than 1% to more than 10%. Related to our pur-

pose, in this study the action of opening a new win-

dow is only registered when it is done from the menu

item. Moreover, this study also mentions that tabs

are used by participants, but no ﬁgure about their

frequency is provided. In (Viermetz et al., 2006), a

clicktree model is proposed, in which all the possible

tabbed paths of user logs are stored into a tree. Using

this model, they found that tabbing is performed by

users from 4% to 85% of the time, which is a quite

large range. Moreover, handling such clicktrees in-

volves huge time and space complexities and is not

appropriate for common web usage mining applica-

tions such as web recommendation.

(Jin et al., 2005) propose a web recommenda-

tion system able to discover task-level patterns. The

From the point of view of parallel browsing, opening a

new window can be considered as being similar to opening

a new tab.

authors use probabilistic latent semantic analysis to

characterize users’ navigational tasks. They then use

a bayesian updating to compute the probability of

each task being performed according to a given active

session. Then recommendations are computed using

a maximum entropy model in which one of the fea-

tures uses a ﬁrst-order Markov model. So, the result-

ing model can detect task changes within one session

and may be considered as a viable model for parallel

browsing data, as it implicitly recognizes the differ-

ent tasks. In the same spirit, (Bonnin et al., 2009)

propose a skipping-based recommender that implic-

itly takes into account parallel browsing. This model

is based on a discontiguous version of Markov models

and computes recommendations based on a weighted

combination of subhistories, i.e. small subsequences

of user’s current session. In this paper we go a step

farther and identify more explicitly which resources

in a user session correspond to which tasks.

3 EXTRACTING LINEAR

SESSIONS

We now focus on the detection of tasks within ses-

sions. We ﬁrst deﬁne several concepts:

• We call a task a typical sequence of resources that

can have several slight variations;

• We call linear session a session in which only

one task is performed. Several different linear

sessions may correspond to the same task while

one linear session cannot correspond to different

tasks;

• We call nonlinear session a session in which sev-

eral tasks are mixed;

• Given two observed sessions X and Y, we say that

X is a subsession of Y if X is included in Y;

• We deﬁne an interrupted session as a linear ses-

sion that was abruptly stopped.

We now propose a new algorithm to extract linear

sessions. A ﬁrst solution would be to make the as-

sumption that if a session X is a subsession of another

session Y, then Y is not a linear session. However,

this way to do is not sufﬁcient as X may simply corre-

spond to an interrupted session. Thus, the assumption

on which is based our proposition for the extraction of

linear sessions is the following: if two sessions X and

Y do not correspond to the same task, and are both

subsessions of a third session Z, then Z is not a linear

session.

To extract linear sessions, a comparison of all the

sessions is performed, and the corresponding non-

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

506

linear sessions are suppressed whenever the afore-

mentioned situation is encountered. The correspond-

ing algorithm is detailed in Algorithm 1. This ex-

traction algorithm lies on two additional algorithms:

sameTask(X,Y) that indicates whether two sessions

X and Y are similar enough to be considered as be-

longing to the same task, and isSubsession(X,Y) that

indicates whether X is a subsession of Y. These algo-

rithms are deﬁned in the following sections.

Algorithm 1: Extraction of the linear sessions.

Data: a list S of sessions

Result: A sublist of S with only linear sessions

for each session X in S do

for each session Y after X in the list S do

if NOT sameTask(X,Y) then

for each session Z in S, Z 6= X and Z 6= Y do

if isSubsession(X, Z) and isSubsession(Y, Z)

then remove Z from S;

end

3.1 Discriminating between Tasks

We propose to use a global alignment algorithm, as

used for instance in (G¨und¨uz and

Ozsu, 2003). The

algorithm used by Gnduz and zsu, the FastLSA algo-

rithm, is an enhancement in terms of storage space

of the Needleman-Wunsch algorithm, which is one

of the standard global alignment algorithms. How-

ever, the time complexity of the FastLSA algorithm

is higher than the standard Needleman-Wunsch algo-

rithm. Storage space is not problematic when com-

puting the best alignment of two sessions, as the size

of a session is usually small. We thus used the stan-

dard Needleman-Wunsch algorithm. Further details

on this algorithm can be obtained in (Needleman and

Wunsch, 1970).

Then, X is considered as belonging to same task

as Y if their global alignment score value is close to

the size of X (or equivalently Y).

3.2 Identifying Subsessions

Another element of our linear session extraction algo-

rithm is how to determine whether a givensession X is

a subsession of a given session Y. A ﬁrst way to deter-

mine this would be to simply check whether each ele-

ment of X can be found in Y in the same order. How-

ever, as mentioned in the previous section, the ses-

sions associated with one given task may show slight

differences, and using such a strategy would be too

restrictive. We thus propose to use an adapted local

alignment algorithm, based on the Smith-Waterman

algorithm (Smith and Waterman, 1981). We chose

the Smith-Waterman algorithm because it is one of

the standard local alignment algorithms.

The classical local alignment problem aims at

ﬁnding the best alignment between two arbitrary sub-

sequences of the input sequences X and Y. Our aim

in this paper is a little bit different. Indeed, our goal

is to detect overlapping of subsessions. The problem

is thus to determine whether X can be aligned with

a discontiguous subsequence of Y in such a way that

any number of insertions can be applied between the

elements of X without penalizing the ﬁnal alignment

score. We propose to apply a simple modiﬁcation to

the Smith-Waterman algorithm by penalizing only the

insertions between the elements of Y, while not de-

creasing the score when insertions are performed be-

tween the elements of X. The resulting algorithm, de-

tailed in Algorithm 2, is asymmetric. As a result, X

can be considered as being a subsession of Y if the

local alignment score value is close to the size of X.

Algorithm 2: Modiﬁed version of the Smith-Waterman al-

gorithm.

Data: two sequences X and Y of sizes m and n

Result: score β of the optimal alignment

set F(i, 0) = 0 for all i = 0, 1, ..., m;

set F(0, j) = 0 for all j = 1, 2, ..., n;

β ← 0;

for i = 1 to m do

for j = i to n do

F(i, j) ← max











F(i− 1, j − 1) + s(x

, y

)

F(i− 1, j) − d

F(i, j− 1)

;

if F(i, j) > β then β ← F(i, j)

end

We propose the match function and mismatch

penalties to have to following values: d = −1 and

s(x

, y

) = 1 if x

= y

, and −2 else. In this fashion,

if an insertion is performed on X and then another is

performed on Y, the penalty is the same as for a mis-

match (substitution).

4 THE TABAKO

RECOMMENDER

We now present our Tabbing-based recommender.

As mentioned in section 2, the all-k

-order Markov

model is one of the best performing model of the

state-of-the-art. However, one of its major draw-

back is that it only takes into account contiguous se-

DETECTING PARALLEL BROWSING TO IMPROVE WEB PREDICTIVE MODELING

507

quences, and cannot handle efﬁciently parallel brows-

ing and navigation mistakes. Nevertheless, when ap-

plied to linear sessions, itis the most appropriatealgo-

rithm. Indeed, it provides a good accuracy addition-

ally to a full coverage, and a lower time complexity

compared to the other state-of-the art models, espe-

cially open sequential patterns. So, we assume that if

we are able to detect the tasks that have been mixed

within a session and ignore navigation mistakes, then

the all-k

-order Markov model will still be applied to

linear sessions, and thus will provide a good accuracy.

We propose to exploit the algorithms presented in

the previous section in order to extract the tasks that

have been mixed, and then apply the traditional all-

-order Markov model on these tasks. The resulting

recommendation model is the TAbbing-Based All-

-Order Markov model (TABAKO model). Com-

pared to the classical all-k

-order Markov model, two

major enhancements are provided:

During the Training Step. The linear sessions of the

training corpus are ﬁrst extracted and stored in order

to be used during the recommendation step (section

3). Second, the model is built exactly as an all-k

order Markov model. Thus, instead of using all the

training data, only the linear sessions are used. In this

fashion, only useful sequences are used to train the

model;

During the Recommendation Step. Before look-

ing for a matching between the active user session

and one of the states stored in the Markov model, the

best overlapping of linear subsessions in the session

is computed. In this way, the resources of the ses-

sion that correspond to the same task are put together

and theses tasks can be matched to the all-k

-order

Markov model.

The recommendation process is performed in two

steps: (1) the extraction of the best overlapping and

the extraction of the corresponding linear sessions,

and (2) the creation of the recommendation lists by

applying the all-k

-order Markov model on the re-

trieved linear subsessions. The overall recommenda-

tion algorithm is presented in Algorithm 3.

Once the best overlapping has been computed, it

is possible that some resources in the active session

remain unlinked to any of the linear sessions. We then

consider that these resources correspondto navigation

mistakes and we propose to discard them. In this way,

both parallel browsing and noise are handled.

4.1 Extracting the Best Overlapping

In the following, the list of resources that the user has

previously browsed in the session will be referred to

as the history h = hr

, ..., r

i−1

i. The extraction of the

Algorithm 3: Recommendation process.

Data: Current user history h

An all-k

-order Markov model M

A list of linear sessions L

Result: A recommendation list R

subhistories ←

c ←bestSubsession(h,L);

while khk 6= 0 and c 6=

0 do

add c to subhistories;

h ← h− c;

c ←bestSubsession(h,L);

end

for each subhistory c in subhistories do

R ← merge(R, buildList(c, M));

end

best overlapping is performed according to an itera-

tive process:

1. Find the best subsession c of current user history;

2. Remove the matching elements of c in the history;

3. Reiterate on the remaining history, until no re-

source remains or no additional subsession can be

found.

The search of the best subsession of a given user’s

history h is performed using the Smith-Waterman al-

gorithm. For each linear session l extracted and

stored during the training step, all the preﬁxes of l

of size varying from max(klk − 1, khk) to 1 are ﬁrst

extracted. Indeed, as the current user session is still

open, it cannot be compared to the entire linear ses-

sions. For instance, a user may have browsed 4 re-

sources of a given task with a usual length of 10. Last,

the subsession with the highest score is returned and

added to the set of sub-histories. This process is de-

tailed in Algorithm 4.

Algorithm 4: Determining the best subsession.

Data: Current user history h

A list of linear sessions L

Result: The subsession that matches h the best

C ←

for each linear session l in L do

for i ← max(klk − 1, khk) downto 1 do

←substring(l, i);

(c, β) ←Smith-Waterman(l

, h);

add (c, β) to L;

end

return arg max L

4.2 Building the Recommendation Lists

After the determination of the best overlapping, the

corresponding list of subsessions is extracted from the

KDIR 2010 - International Conference on Knowledge Discovery and Information Retrieval

508

user’s active session. Then, for each subsession, a rec-

ommendation list is computed using the all-k

-order

Markov model. The resulting sublists are then merged

according to the following policy: the resources is-

sued from the subsession that has the best alignment

score are put on the top of the list. The other recom-

mendations are then appended to the end of the list,

except when they are already in the list.

5 CONCLUSIONS AND FUTURE

WORK

This paper focused on parallel browsing in the frame

of web predictive models in order to improve rec-

ommendation accuracy. We proposed algorithms to

discriminate between linear and nonlinear sessions.

These algorithms exploit sequence alignment: global

and local alignments. We then proposed a new rec-

ommendation model called TABAKO. This model is

based on an all-k

-order Markov model and exploits

the sequence alignment algorithms to extract linear

sessions from the active user’s session.

In a future work, we plan to validate this position

paper through experiments on a browsing dataset in

which tab switching is performed. Such a study can

include analysis of the amount of linear sessions ob-

tained when using different parameters on our model,

and a comparison to state-the-art models, such as the

all-k

-order Markov model and the approaches for

obtaining more granularity presented in section 2.

REFERENCES

Agrawal, R., Imielinski, T., and Swami, A. (1993). Min-

ing Association Rules between Sets of Items in Large

Databases. In Buneman, P. and Jajodia, S., editors,

Proceedings of the 1993 ACM SIGMOD International

Conference on Management of Data, pages 207–216.

Agrawal, R. and Srikant, R. (1995). Mining Sequential Pat-

terns. In ICDE’95: Proceedings of the International

Conference on Data Engineering, pages 3–14.

Ayres, J., Flannick, J., Gehrke, J., and Yiu, T. (2002). Se-

quential Pattern Mining Using a Bitmap Representa-

tion. In KDD ’02: Proceedings of the eighth ACM

SIGKDD international conference on Knowledge dis-

covery and data mining, pages 429–435.

Bonnin, G., Brun, A., and Boyer, A. (2009). A Low-

Order Markov Model integrating Long-Distance His-

tories for Collaborative Recommender Systems. In

Proceedings of the 13th International Conference on

Intelligent user interfaces (IUI), pages 57–66.

Chi, E., Pitkow, J., Mackinlay, J., Pirolli, P., Gossweiler, R.,

and Card, S. (1998). Visualizing the Evolution of Web

Ecologies. In CHI ’98: Proceedings of the SIGCHI

conference on Human factors in computing systems,

pages 400–407.

Deshpande, M. and Karypis, G. (2004). Selective Markov

Models for Predicting Web Page Accesses. ACM

Trans. Internet Technol., 4(2):163–184.

G¨und¨uz, S. and

Ozsu, M. (2003). A Web Page Prediction

Model Based on Click-Stream Tree Representation of

User Behavior. In KDD ’03: Proceedings of the ninth

ACM SIGKDD international conference on Knowl-

edge discovery and data mining, pages 535–540.

Jianyong, W., Jiawei, H., and Li, C. (2007). Frequent

Closed Sequence Mining without Candidate Mainte-

nance. IEEE Transactions on Knowledge and Data

Engineering, 19(8):1042–1056.

Jin, X., Zhou, Y., and Mobasher, B. (2005). Task-oriented

Web User Modeling for Recommendation. In Pro-

ceedings of the 10th International Conference on User

Modeling (UM), pages 109–118.

Lu, L., Dunham, M., and Meng, Y. (2005). Mining Signif-

icant Usage Patterns from Clickstream Data. In 7th

International Workshop on Knowledge Discovery on

the Web, pages 1–17.

Nakagawa, M. and Mobasher, B. (2003). Impact of Site

Characteristics on Recommendation Models Based

On Association Rules and Sequental Patterns. In In-

telligent Techniques for Web Personalization.

Needleman, S. and Wunsch, C. (1970). A general method

applicable to the search for similarities in the amino

acid sequence of two proteins. Journal of molecular

biology, 48(3):443–453.

Pitkow, J. and Pirolli, P. (1999). Mining Longest Repeat-

ing Subsequences to Predict World Wide Web Surf-

ing. In USITS’99: Proceedings of the 2nd conference

on USENIX Symposium on Internet Technologies and

Systems, pages 139–150.

Schechter, S., Krishnan, M., and Smith, M. (1998). Using

Path Proﬁles to Predict HTTP Requests. Computer

Networks and ISDN Systems, 30(1-7):457–467.

Smith, T. and Waterman, M. (1981). Identiﬁcation Of Com-

mon Molecular Subsequences.

Srivastava, J., Cooley, R., Deshpande, M., and Tan, P.

(2000). Web Usage Mining: Discovery and Appli-

cations of Usage Patterns from Web Data. SIGKDD

Explorations Newsletter, 1(2):12–23.

Tan, P. and Kumar, V. (2002). Discovery of Web Robot

Sessions Based on their Navigational Patterns. Data

Mining Knowledge Discovery, 6(1):9–35.

Viermetz, M., Stolz, C., Gedov, V., and Skubacz, M. (2006).

Relevance and Impact of Tabbed Browsing Behavior

on Web Usage Mining. In Proceedings of the 2006

IEEE/WIC/ACM International Conference on Web In-

telligence (WI), pages 262–269.

Weinreich, H., Obendorf, H., Herder, E., and Mayer, M.

(2006). Off the Beaten Tracks: Exploring Three As-

pects of Web Navigation. In WWW ’06: Proceedings

of the 15th international conference on World Wide

Web, pages 133–142.

DETECTING PARALLEL BROWSING TO IMPROVE WEB PREDICTIVE MODELING

509