Sistema de recomendação baseado em agrupamento usando Propagação de Afinidades

Abstract

Recommend items based on similarity of interests (Collaborative Filtering) is attractive to many domains: books, movies, music, products and etc. However, it's not always works well due to the fact of collections of items as scattered as in companies such as Amazon, Netflix, Spotify, among others. Clustering based collaborative filtering proposes greater scalability for very sparse collections, its premise is if the person a and person b like the same set of movies, then the person a probably will like other movies the person b likes. Clustering people into groups based on the items they bought, one can get good recommendations for items to be bought, in this way, predictions can be made by crowding people into groups, based on the movies they watched(user-based) and/or groups of movies which tend to be of the taste of the people the same interest(item-based). The K-means is a classic clustering algorithm, being simple, efficient and widely used, however, it comes with some restrictions: the number of final groups must be defined a priori, very sensitive to the initial choice of centroids in the creation of groups, it can generate empty groups, among others. The algorithm Affinity Propagation is an alternative to K-means, it is a recently proposed algorithm that has gained great popularity in areas of bioinformatics, presen-ting good results for problems like clustering DNA sequences, and being applied also for clustering of faces (image), collections of films and summarization of texts. The document presents an approach of Recommender Systems based on clustering using Affinity Propagation in order to investigate whether the good results that the algorithm has in other areas are also valid for recommender systems area.