Temporal algorithms for collaborative filtering
DEI - aula 3 B
3 febbraio 2011
Recommender systems help users in distinguishing interesting products from irrelevant ones. A large number of recommender algorithms are based on collaborative filtering: such methods usually rely on a matrix containing the ratings that the users have assigned to some of the items, expressing the extent of their preference. Several datasets associate the ratings with the time at which they have been provided, but the use of such temporal information has been neglected for a long time. A recent work has shown that temporal algorithms can improve collaborative filtering performance, since the preferences of the users change over time as well as their personal rating scales; even the identity of the person accessing an account may vary. As a consequence, in order to recommend a user at a certain time instant, ratings provided close to that instant represent a more reliable information. Despite the exciting performance, the algorithms proposed in the literature require a heavy training phase, with the tuning of s everal learning constants.
In this work more usable and manageable temporal algorithms are proposed; in more detail, existing methods not requiring a massive training are enriched with temporal features. The proposed approaches are evaluated with accuracy metrics rather than with the more usual error ones on well-known datasets, showing encouraging improvements with respect to the static algorithms. A further set of experiments is carried out to assess the ability of the described strategies in recommending unpopular but relevant items.
Area di ricerca:
Web, multimedia e database