An Improved Neighborhood-Aware Unified Probabilistic Matrix Factorization Recommendation

Abstract

Recommendation systems require sufficient information to provide proper recommendations. Both rating and tagging information can be used in social tagging systems. Many recommendation systems consider the relationships between users, items and tags, which affect the recommendation results. To address this issue, this paper proposes a neighborhood-aware unified probabilistic matrix factorization recommendation model that fuses social tagging. In the proposed approach, the similarities between users and items are first calculated by using tags to make neighborhood selections. Then, a user–item rating matrix, a user–tag tagging matrix, an item–tag correlation matrix and a unified probabilistic matrix factorization are constructed to obtain the latent feature vectors of three matrices to be recommended to users by optimizing the training parameters. In the experiments, the proposed model is compared with three other collaborative filtering approaches on the MovieLens dataset to evaluate its performance. The experimental results demonstrate that the proposed model uses the tag semantics effectively and improves the recommendation quality.