CLUSTERING Y DIVERSIDAD EN SISTEMAS DE RECOMENDACIÓN TOP-N

Abstract

Recommender Systems aim to help people dealing with information overload. Collaborativefiltering (CF) is one of the most successful techniques used in recommendersystems. It is based on the idea that people often get the best recommendationsfrom someone with similar tastes to themselves. Broadly, there are model-based andmemory-based CF techniques. The former, learn a model to make predictions. The latter,uses similarity measures to compute the proximity between users (User-based) oritems (Item-based) and build a neighborhood (Neighborhood-based CF). UBCF, whileeffective, suffers from scalability problems as the database grows. To address the scalabilityissue, clustering-based CF algorithms constraint the seek of users within smalluser clusters instead of the entire database. However, there is a trade-off between efficiencyand prediction accuracy.In this Msc. Thesis, we present a novel approach that combines the advantages ofUBCF and cluster-based CF methods by introducing a cluster-based distance functionused for neighborhood computation. To expand the search of relevant users/items weuse a novel measure that is able to exploit the global cluster structure to infer user’sdistances. Empirical studies on widely known benchmark datasets suggest that our proposalis feasible. Nevertheless, recommender systems are frequently evaluated usingindexes based on variants and extensions of precision-like measures. As these measuresare biased toward popular items, a list of recommendations just need to include a fewpopular items to perform well. To provide a more robust and realistic evaluation of ourproposed method, in the second part of this Thesis, new approaches for novelty anddiversity evaluation have been proposed. Experimental results show that our proposedmethod, based on cluster models, can promote diversity retrieval.