Section: New Results

Recommender Systems

A Biclustering Approach to Recommender Systems

Participants : Florestan de Moor, Davide Frey.

Recommendation systems are a core component of many e-commerce industries and online services since they ease the discovery of relevant products. Because catalogs are huge, it is impossible for an individual to manually search for an item of interest, hence the need for some automatic filtering process. Many approaches exist, from content-based ones to collaborative filtering that include neighborhood and model-based techniques. Despite these intensive research activities, numerous challenges remain to be addressed, particularly under real-time settings or regarding privacy concerns, which motivates further work in this area. We focus on techniques that rely on biclustering, which consists in simultaneously building clusters over the two dimensions of a data matrix. Although it was little considered by the recommendation system community, it is a well-known technique in other domains such as genomics. In work [42] we present the different biclustering-based approaches that were explored. We then are the first to perform an extensive experimental evaluation to compare these approaches with one another, but also with the current state-of-the-art techniques from the recommender field. Existing evaluations are often restrained to a few algorithms and consider only a limited set of metrics. We then expose a few ideas to improve existing approaches and address the current challenges in the design of highly efficient recommendation algorithms, along with some preliminary results.

This work was done in collaboration with Antonio Mucherino (University of Rennes 1).

Unified and Scalable Incremental Recommenders with Consumed Item Packs

Participant : Erwan Le Merrer.

Recommenders personalize the web content by typically using collaborative filtering to relate users (or items) based on explicit feedback, e.g., ratings. The difficulty of collecting this feedback has recently motivated to consider implicit feedback (e.g., item consumption along with the corresponding time). In this work [39], we introduce the notion of consumed itempack (CIP) which enables to link users (or items) based on their implicit analogous consumption behavior. Our proposal is generic, and we show that it captures three novel implicit recommenders: a user-based (CIP-U), an item-based (CIP-I),and a word embedding-based (DEEPCIP), as well as a state-of-art technique using implicit feedback (FISM). We show that our recommenders handle incremental updates incorporating freshly consumed items. We demonstrate that all three recommenders provide a recommendation quality that is competitive with state-of-the-art ones, including one incorporating both explicit and implicit feedback

This work was done in collaboration with Rachid Guerraoui (EPFL), Rhicheek Patra (Oracle) and Jean-Ronan Vigouroux (Technicolor).