Factored Item Similarity and Bayesian Personalized Ranking for Recommendation with Implicit Feedback

Abstract

Item recommendations aim to predict a list of items (e.g., items on Amazon website) for each user that he or she might like. In fact, implicit feedback, such as transaction records in e-commerce websites and the “likes” behavior in social networks website (e.g., Facebook), has been received more and more attention in the scenarios of item recommendation. The core of the recommender system is the ranking algorithm which exploits the implicit feedback and generates the personalized item list to meet user’s specific preferences. In most of the previous studies, the pairwise personalized ranking techniques empirically achieve better performance than the matrix factorization and adaptive k nearest-neighbor method since the pairwise ranking methods can directly reflect the model user’s ranking preference on items. In most of the recent works, factored item similarity techniques which learn the global item similarity by utilizing two low-dimensional latent factor matrices achieve better performance than other state-of-art top-N methods with predefined similarity, such as cosine similarity. The individual relative preference assumption among observed items and unobserved items are critical for the pairwise ranking methods. As a response, this paper proposes a new and improved preference assumption based on the factored item similarity and individual preference. In addition, a novel recommendation algorithm correspondingly named factored item similarity and Bayesian Personalized Ranking model is designed. The novelty of the algorithm is that it can (1) learn the global item similarity with latent factor models. (2) utilize effective pairwise ranking methods to deal with the item recommendation problems with implicit feedback. (3) assign different item weights on explicit feedback and implicit feedback. Empirical results show that this model outperforms other state-of-the-art top-N recommendation methods on two public datasets in terms of prec@5 and ndcg@5. It can be found that the advantage of FSBPR lies in its ability to exploit implicit feedback and capture global item similarity.