Thèse soutenue

Evaluation des systèmes de recommandation à partir d'historiques de données

FR
Auteur / Autrice : Bruno Pradel
Direction : Patrick Gallinari
Type : Thèse de doctorat
Discipline(s) : Informatique
Date : Soutenance en 2013
Etablissement(s) : Paris 6

Mots clés

FR

Mots clés contrôlés

Résumé

FR  |  
EN

This thesis presents various experimental protocols leading to abetter offline estimation of errors in recommender systems. As a first contribution, results form a case study of a recommendersystem based on purchased data will be presented. Recommending itemsis a complex task that has been mainly studied considering solelyratings data. In this study, we put the stress on predicting thepurchase a customer will make rather than the rating he will assign toan item. While ratings data are not available for many industries andpurchases data widely used, very few studies considered purchasesdata. In that setting, we compare the performances of variouscollaborative filtering models from the litterature. We notably showthat some changes the training and testing phases, and theintroduction of contextual information lead to major changes of therelative perfomances of algorithms. The following contributions will focus on the study of ratings data. Asecond contribution will present our participation to the Challenge onContext-Aware Movie Recommendation. This challenge provides two majorchanges in the standard ratings prediction protocol: models areevaluated conisdering ratings metrics and tested on two specificsperiod of the year: Christmas and Oscars. We provides personnalizedrecommendation modeling the short-term evolution of the popularitiesof movies. Finally, we study the impact of the observation process of ratings onranking evaluation metrics. Users choose the items they want to rateand, as a result, ratings on items are not observed at random. First,some items receive a lot more ratings than others and secondly, highratings are more likely to be oberved than poor ones because usersmainly rate the items they likes. We propose a formal analysis ofthese effects on evaluation metrics and experiments on the Yahoo!Musicdataset, gathering standard and randomly collected ratings. We showthat considering missing ratings as negative during training phaseleads to good performances on the TopK task, but these performancescan be misleading favoring methods modeling the popularities of itemsmore than the real tastes of users.