Abstrakt
In many application domains of recommender systems, content-based information are available for users, objects or both. Such information can be processed during recommendation and significantly decrease the cold-start problem.
However, content information may come from several, possibly external, sources. Some sources may be incomplete, less reliable or less relevant for the purpose of recommendation.
Thus, each content source or attribute possess different level of informativeness, which should be taken into consideration during the process of recommendation. In this paper, we propose a multiple content alignments extension to the Bayesian Personalized Ranking Matrix Factorization (BPR-MCA).
The proposed method incorporates multiple sources of content information in the form of user-to-user or object-to-object similarity matrices and aligns users' and items' latent factors ac-cording to these similarities. During the training phase, BPR-MCA also learns the relevance weight of each similarity matrix.
BPR-MCA was evaluated on the MovieLens 1M dataset, extended by the content information from IMDB, DBTropes and ZIP code statistics. The experiment shows that BPR-MCA can help to significantly improve recommendation w.r.t. nDCG and AUPR over standard BPR under several cold-start scenarios.