Joint latent factors and attributes to discover interpretable preferences in recommendation

Abstract

Latent factor model (LFM), which uses a dot product between the resulting user and item latent factors to rank candidate items, is the most popular collaborative filtering (CF) based method in recommender systems, due to its better generalization and requiring little expert knowledge. However, LFM suffers from cold-start problems in which no historical interactions are available for new users or items and explanation problems in which no clear explanations of the recommended items are provided for users. Meanwhile, side information like user attributes (e.g. gender, age, occupation, etc.) and item attributes (e.g. category, price, popularity, comments, etc.) is usually rich in a recommender system. To better cope with the cold-start and explanation problems in LFM at the same time, we propose a method which combines latent factors and attribute information to discover interpretable preferences. First, we adapt the sampling strategy of pairwise learning in Bayesian personalized ranking - matrix factorization (BPR-MF) to incorporate explicit feedback into our rating-aware framework. Next, latent factors, preference and attribute factors are fused into the user-item affinity scores’ computation. Therefore, collaborative and content information are better utilized in a hybrid method. In this way our method can effectively improve the recommendation performance under conventional recommender settings. Furthermore, the devised mapping functions learn how to map attribute factors to latent, preference and bias factors, which can deal with the cold-start problems, yielding accurate and fast recommendation results. Besides, the learned preference factors are interpretable. The preference-based explanations for recommendation results can be generated automatically to improve user experience. Our algorithm is evaluated on several large-scale datasets with comparisons to state-of-the-art methods. The experimental results demonstrate that our approach is more effective for recommendation, as well as dealing with cold-start scenarios. Moreover, the results also show that our method can provide satisfactory explanations.