Tempweb13

Discovering temporal hidden contexts in web sessionsfor user trail prediction

Julia Kiseleva (Eindhoven University of Technology),Hoang Thanh Lam (Eindhoven University of Technology),Mykola Pechenizkiy (Eindhoven University of Technology),Toon Calders (Université Libre de Bruxelles)

User intention prediction

Search Refine Search

PaymentClick Product

ViewClick ?

What is next?

• Does exist any contextual information?• How we can discover it?• How we can utilize it?

Contextual Partitioning

• Approaches to create local models:o Horizontal partitions

Users from Europe

Users from South America

Session 1 Search

Refine Search

Click on Banner

Product View

Payment

Session 3 Product View

Payment

Session 3 Search

Refine Search

Refine Search

Click on Banner

Session 4 Search

Refine Search

Click on Banner

Product View

Payment

Session 5 Product View

Click on Banner

Search

User navigation graph

Search

Refine Search

Payment

Click on Banner

Product View

1.0 2/3

1/3

1/21/4

Drop out

3/4

1/4

1

1/4

Horizontal Partitioning

• Approaches to create local models:o Horizontal partitiono Vertical partition :

• Two types of behavior:o Ready to buy – (Product View, Payment)

Session 1

Search Refine Search

Click on Banner

Product View

Payment

Session 2

Product View

Payment

Session 3

Search Refine Search

Refine Search

Click on Banner

Session 4

Search Refine Search

Click on Banner

Product View

Payment

Session 5

Product View

Click on Banner

Search

Contextual Partitioning

• Approaches to create local models:o Horizontal partitiono This work is about Vertical partition:

• Two types of behavior:o Ready to by – (Product View, Payment)o Just browsing – (Search, Refine Search, Click on Banner)

Session 1

Search Refine Search

Click on Banner

Product View

Payment

Session 2

Product View

Payment

Session 3

Search Refine Search

Refine Search

Click on Banner

Session 4

Search Refine Search

Click on Banner

Product View

Payment

Session 5

Product View

Click on Banner

Search

Contextual Partitioning

• Approaches to create local models:o Horizontal partitiono Vertical partition:

• Two types of behavior:o Ready to buy – (Product View, Payment)o Just browsing – (Search, Refine Search, Click on Banner)

Session 1

Search Refine Search

Click on Banner

Product View

Payment

Session 2

Product View

Payment

Session 3

Search Refine Search

Refine Search

Click on Banner

Session 4

Search Refine Search

Click on Banner

Product View

Payment

Session 5

Product View

Click on Banner

Search

Contextual Partitioning

Problem Description

Timeline

Search Refine Search

PaymentClick Product

ViewSearch Click

Context ``Find information”

Context ``Buy product”

What is next?Probably user will change

intent?

Problem Description

Timeline

Search=A

Refine Search

=B

PaymentClick

=C

Product View

=D

Search Click

Context ``Find information”

Context ``Buy product”

Context-Awareness Example

Contextual Feature:User intent

DATA

Contextual Categories

Individual Models

Mapping G

Mapping H Context

Discovery

Ready to buy

Just browsing

Context-Awareness

………………

Contextual features

DATA Environment

Contextual Categories

Individual Learners

Mapping G

Mapping H

Context Discovery

Context-Awareness: (G,H)

Temporal Context-Awareness

G H

Temporal Context-Awareness: (G,H,ti)

……

..

G

G

H

H

Web

Sessio

n

S

Contextual features

Contextual Categories

Individual Models

Discovery hidden contexts

Web log

Train:To train

predictive models

Validation:To find Best

clusters

Test:To derive final

accuracy

To find a “Best”

clusters

Calculate final accuracy

To train local predictive

models

Optimization problem

a b c d abababababcdcdababcdcdcd

The number of true predictions = 0

a b c d

1.0 1.0 1.0

1.0

Hierarchical clustering

a b c d

ab

abababababcdcdababcdcdcd

The number of true predictions = 12

a b c d

1.0

1.0

1.0 1.0

Hierarchical clustering

a b c d

ab

abababababcdcdababcdcdcd

The number of true predictions = 20

a b c d

1.0

1.0

1.0

1.0

cd

Hierarchical clustering

a b c d

ab

abababababcdcdababcdcdcd

The number of true predictions = 20

a b c d

1.0

1.0

1.0

1.0

cd

abcd

Stop as long as no additional prediction benefit of merging

Hierarchical clustering

Mastersportal.eu - Homepage

Quick Search

Banner Click

Universities in the spotlight

Mastersportal.eu - Search

Refine Search

Click on Program is Search Result

Click on University

Click on Country

DatasetDate May 2012

#sessions

350.618

#requests

1.775.711

#sessions from Eu

159.991

Results

Explaining the results

eb C d

a1.0

1.0

0.9

0.1

1.0

1.0

Explaining the results

eb C d

a1.0

1.0

0.9

0.1

1.0

1.0

Explanation:1. Users’ behavior doesn’t follow Markov

property2. Ambiguous Transition Matrix

Explaining the results

University on nearby universities

Related Programs

0.315

0.299

Quick Search

0.306

Related Programs

Related Programs

…

Click on Program

0.283

…

Resulted ClustersId Summary Cluster

1 Intensive Search Basic Search, Refine Search, EmptySearch Result

2 Explore informationrelated toprogram

Program impression in search result,Banner click, Program click ,Click on university link

3 Start ofbrowsing

University Spotlight impression,Quick search

4 Explorecountry information

File view, Click on country link

5 Exploresearch result

Program impressions in search results,University impression onnearby universities

6 Explore program Program in landing page, Submitinquiry

7 Outlier Submit question, X-node

Site Map• Page is represented as set of possible actions

o Example: Homepage is (Quick Search, University Spot light impression, Question Submit)

o Calculate Jaccard similarity between Page and Cluster

Site map

Conclusion• We formulate the problem of temporal context

discovery as an optimization problem.

• A hierarchical clustering method is proposed to determine the optimal number of hidden contexts and mine temporal contexts.

• We show a real-world use case in which the contexts as we defined them do exist and are useful for prediction.

Future works• Testing our method on another datasets

• Introducing a mechanism for detecting context switching within a web-session

• Considering multidimensional contextualfeatures.

Thank you!• Context identification and integration

it into prediction models• Accurately predicting users’ desired

actions and understanding behavioral patterns of users in various web-applications

• Personalization and adaptation to diverse customer need and preferences

• Accounting for the practical needs within the considered application areas.

Conclusion• We formulate the problem of temporal context

discovery as an optimization problem.

• A hierarchical clustering method is proposed to determine the optimal number of hidden contexts and mine temporal contexts.

• We show a real-world use case in which the contexts as we defined them do exist and are useful for prediction.

Questions?

Outline• Problem description• Def. context-aw => temporal CA• Motivational example• Objectives• Method: HC• Data• Results• Summary• Conclusion• Future works

• Web sessions:• Type of events:• Web session is an

ordered sequence of events:

• Space of contextual features:

• Predictive model:• Evaluation function:• Contextual categories: