Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

26
RecSys 2015 – 2nd Workshop on New Trends in Content-Based Recommender Systems September 16 - 20, 2015 in Vienna, Austria Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity Paolo Tomeo, Tommaso Di Noia, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro, Eugenio Di Sciascio {paolo.tomeo, tommaso.dinoia, eugenio.disciascio}@poliba.it {marco.degemmis, pasquale.lops, giovanni.semeraro}@uniba.it Polytechnic University of Bari - Bari (ITALY) University of Bari Aldo Moro - Bari (ITAL

Transcript of Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Page 1: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

RecSys 2015 – 2nd Workshop on New Trends in Content-Based Recommender SystemsSeptember 16 - 20, 2015 in Vienna, Austria

Exploiting Regression Trees as User Models forIntent-Aware Multi-attribute Diversity

Paolo Tomeo, Tommaso Di Noia, Marco de Gemmis, Pasquale Lops, Giovanni Semeraro, Eugenio Di Sciascio

{paolo.tomeo, tommaso.dinoia, eugenio.disciascio}@poliba.it {marco.degemmis, pasquale.lops, giovanni.semeraro}@uniba.it

Polytechnic University of Bari - Bari (ITALY) University of Bari Aldo Moro - Bari (ITALY)

Page 2: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Introduction

ProblemHow to diversify with different items attributes

complying with the users interests?

ProposalRegression trees to represent user interests as a combination of characteristics

Evaluation in terms of individual and aggregate diversity

Page 3: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Individual Diversity

Same Artist – Low Diversity Diverse Artists

Page 4: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Diversity and Satisfaction

Recommendations quality requires diversity even at the cost of reducing accuracy [Ziegler at al. WWW ‘05, McNee et al. CHI ‘06, Zhou at al. PNAS ‘10, Adamopoulos and Tuzhilin RecSys ‘11, Hurley and Zhang TOIT ‘11, Vargas at al. RecSys ‘14, …]

Diversity has a significant positive influence on the user satisfaction (study with 500+ users of Movielens) [Ekstrand et al. “User Perception of Differences in Recommender Algorithms”, RecSys ‘14]

Page 5: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

How to Diversify

Data

Top-M List

Re-Ranked Top-N List

Recommender System

DiversificationAlgorithm

Items description

User profile

M >> N

Page 6: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Diversification Algorithms

Several algorithms presented so far: - MMR (Carbonell and J. Goldstein, SIGIR ‘98)

- IA-Select (Agrawal et al., WSDM ‘09)

- xQuAD (Santos et. al, WWW ’10)- …- BinomDiv (Vargas et. al RecSys ‘14)

Need of information about items- content- external information- statistical information- …

Should take user interests into consideration (intent-aware)- user model

Page 7: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Greedy strategy

Page 8: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

xQuAD

eXplicit Query Aspect Diversification

likelihood of item i being chosen given the feature f

the user u interest in the feature f

Page 9: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

xQuAD

eXplicit Query Aspect Diversification

Penalization of redundancy

Page 10: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Multi-Attribute Diversity

Attribute FeaturesDirector Cameron CroweYear of release 2001Actor Tom Cruise, Cameron Diaz,

Penélope Cruz

Genre Fantasy, Mystery, Romance

Attribute FeaturesDirector Cameron CroweYear of release 1996Actor Tom Cruise, Cuba Gooding Jr.,

Renée Zellweger

Genre Comedy, Drama, Romance

Page 11: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Regression Trees as User Models

A regression tree allows to represent user tastes as acombination of interrelated characteristics

Romance Movie

falsetrue

AlPacino acts in

falsetrue

Direct by Crowe

falsetrue

interest = 5 interest = 5 interest = 3Year < 2000

falsetrue

interest = 5 interest = 2.3

Page 12: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Rules from Regression Trees

M5Rules produces rules from regression trees.

Good compromise between rules accuracy and compactness

http://weka.sourceforge.net/doc.dev/weka/classifiers/rules/M5Rules.html

Page 13: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Rules in Diversification

Data

Recommender System

DiversificationAlgorithm

Items description

User profile

Rule1

Rule2

.

.

.Rulen

Rule1 , Rule3

Rule1

Rule1

Rule1 , Rule3

Rule1 , Rule2

Rule1

Rule5 , Rule6

Rule2

Rule3 , Rule4 Rule5 Rule2 , Rule3 Rule1 , Rule7 Rule1

Page 14: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

xQuAD adapted for multi-attribute

Original

Multi-attribute

domain of attribute A

Page 15: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

xQuAD adapted for Rules

Original

For rules

rules of user u matched by item i importance of rule

m for user uimportance of the rule m for item i

Page 16: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

xQuAD adapted for Rules

Original

For rules

RT binary function

DivRT avg similarity between m and each rule covered by item j

Page 17: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Evaluation

Page 18: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Datasets

Mappingshttp://sisinflab.poliba.it/semanticweb/lod/recsys/datasets/

• Movielens 1M mapped with Dbpedia-attributes: Genre, Decade of Release, Actors, Directors

- 998,963 ratings from 6,040 users on 3,625 movies- sparsity 95.7%- split 60-40%

• LibraryThing mapped with Freebase-attributes: Genre, Author, Subject

- 565,310 ratings from 7,278 users on 27,358 books- sparsity 99.7%- split 80-20%

Page 19: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Evaluation Metrics

Accuracy Precision RecallnDCG

Individual Diversity Intra-List Diversity (ILD)redundancy-aware nDCG (α-nDCG)

Aggregate Diversity Catalog coverage Entropy

Page 20: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Evaluation Setting

Two steps

1. Top-200 recommendations computed with BPRMF (MyMediaLite)

2. Top-10 diverse recommendations • repeated varying the value of λ from 0 to 0.95 (step 0.05)

Page 21: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Compared Algorithms

Baseline- xQuAD for multi-attribute

Proposed- RT: xQuAD for Rules- DivRT: xQuAD for Rules with a diversity analysis between rules

Combination- xQuAD-after-RT: xQuAD on top-50 from RT- RT-after-xQuAD: RT on top-50 from xQuAD

Page 22: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Individual diversityMovielens LibraryThing

ILD

α-nDCG

Page 23: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Aggregate diversityMovielens LibraryThing

Coverage

Entropy

Page 24: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Outcome

Rule-based approach - less individual diversity + better aggregate diversity

Combination of baseline and proposed approach+ good compromise among accuracy, individual and aggregate diversity

Page 25: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

ConclusionWe proposed a method to leverage regression trees as user model technique for intent-aware multi-attribute diversity problem

We showed that combining attribute-based and rules-based re-rankings obtains the advantages of both

Future work- propose a method for combining attributes and rules within the same formula - evaluation of impact of our approach on recommendation novelty- evaluation with user study

Page 26: Recsys 2015: Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute Diversity

Exploiting Regression Trees as User Models for Intent-Aware Multi-attribute DiversityP. Tomeo, T. Di Noia, M. de Gemmis, P. Lops, G. Semeraro, E. Di Sciascio

Thanks for your attention!

Q & A