+ All Categories
Home > Technology > Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Date post: 08-Sep-2014
Category:
Upload: antonio-moreno
View: 1,790 times
Download: 0 times
Share this document with a friend
Description:
Presentation at workshop on recommender systems at WI-2014. Automatic learning of keyword-based preferences through the analysis of the implicit information provided by the interaction of the user.
Popular Tags:
25
Dynamic learning of keyword-based preferences for news recommendation A.Moreno, L.Marin, D.Isern, D.Perelló ITAKA-Intelligent Tech. for Advanced Knowledge Acquisition Departament d’Enginyeria Informàtica i Matemàtiques Universitat Rovira i Virgili, Tarragona http://deim.urv.cat/~itaka
Transcript
Page 1: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Dynamic learning of keyword-based preferences for news recommendation

A.Moreno, L.Marin, D.Isern, D.Perelló

ITAKA-Intelligent Tech. for Advanced Knowledge Acquisition

Departament d’Enginyeria Informàtica i Matemàtiques

Universitat Rovira i Virgili, Tarragona

http://deim.urv.cat/~itaka

Page 2: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Outline of the talk

Introduction: motivation of the problem

User profile management

Automatic learning of user interests

Evaluation

Conclusions

Page 3: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Outline of the talk

Introduction: motivation of the problem

User profile management

Automatic learning of user interests

Evaluation

Conclusions

Page 4: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Introduction: preference learning

Important issue in recommender systems:

discover the user interests to provide

accurate recommendations.

User preferences may be explicitly given by

the user or may be inferred through the

analysis of his/her actions.

We focus our attention on the case in which

the objects to be recommended are purely

textual (e.g. News).

Page 5: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)
Page 6: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Outline of the talk

Introduction: motivation of the problem

User profile management

Automatic learning of user interests

Evaluation

Conclusions

Page 7: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Representation of preferences

The user profile will store a dynamic set of

keywords. Each of them will have a

positive/negative level of preference, in the

range [-100, 100]

Manchester United +80

Angela Merkel -90

tennis 0

Page 8: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Representation of a textual object

Given a corpus of textual documents, an

object (news) will be represented by a set of n

relevant keywords, determined by the standard

TF-IDF measure.

Page 9: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Evaluation of a textual object

Given a user profile P and a document d, the

score assigned to the document in the first

ranking phase is the addition of the user

preferences on the document’s keywords

Keywords of the

document

Preference

value of

keyword w

Page 10: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Outline of the talk

Introduction: motivation of the problem

User profile management

Automatic learning of user interests

Evaluation

Conclusions

Page 11: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Selected / Over-ranked alternatives

Over-ranked alternatives

Page 12: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Increase preference value

Page 13: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Smaller increase of preference value

Page 14: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Decrease preference value

Page 15: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Summary of learning algorithm (I)

Increase the preference value of the keywords

of the selected news that do not appear in the

over-ranked alternatives.

The more over-ranked alternatives, the greater the

increase

Increase (in a smaller degree) the preference

value of the keywords of the selected news

that appear in the over-ranked alternatives.

The more repetitions on the over-ranked

alternatives, the smaller the increase.

Page 16: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Summary of learning algorithm (II)

Decrease the preference value of the

keywords of the over-ranked alternatives that

do not appear in the selected news.

The more repetitions on the over-ranked

alternatives, the greater the decrease.

The amounts of increase/decrease were

determined empirically, and the details may be

found in the paper.

Page 17: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Outline of the talk

Introduction: motivation of the problem

User profile management

Automatic learning of user interests

Evaluation

Conclusions

Page 18: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Evaluation framework

Retrieval of 6000 news from The Guardian.

Definition of an ideal profile to be learnt.

Random generation of 10 initial profiles.

A single test consists in a series of 400 recommendations over 6000 alternatives, considering 15 alternatives at each step and 30 keywords/news

After each recommendation, the normalised distance between the current profile P and the ideal one I is calculated

Page 19: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)
Page 20: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)
Page 21: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Outline of the talk

Introduction: motivation of the problem

User profile management

Automatic learning of user interests

Evaluation

Conclusions

Page 22: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Conclusions

User preferences on textual documents may

be efficiently learned in an implicit way if the

user has a frequent interaction with the

system.

In the future work we intend to introduce

semantic information in the learning process

If a user likes tennis/football/golf, the system

could infer a general interest on sports.

Treat natural language phenomena like

synonymity and polysemy.

Page 23: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Dynamic learning of keyword-based preferences for news recommendation

A.Moreno, L.Marin, D.Isern, D.Perelló

ITAKA-Intelligent Tech. for Advanced Knowledge Acquisition

Departament d’Enginyeria Informàtica i Matemàtiques

Universitat Rovira i Virgili, Tarragona

http://deim.urv.cat/~itaka

Page 24: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)
Page 25: Dynamic learning of keyword-based preferences for news recommendation (WI-2014)

Recommended