Deep Latent Factor Model with Hierarchical Similarity Measure for recommender systems

Abstract

Latent Factor Model(LFM), as an effective feature mapping method, is widely applied in recommender systems. One challenge of LFM is previous methods usually use the inner product to calculate the similarity between users and items in the latent space, which cannot characterize different impacts of various latent factors. Another challenge is the performance of LFM will be negatively affected when facing data sparsity problem. In this paper, we propose a model named DLFM-HSM(Deep Latent Factor Model with Hierarchical Similarity Measure) to overcome the challenges above. More specifically, we introduce a hierarchical similarity measure to calculate an impact score which can better represent the similarity between a user and an item than the inner product. Also, in order to ease the data sparsity problem, we extract latent representations of users and items using deep neural networks from items’ content information instead of only from user-item rating records. By representing users with items they purchased, our model guarantees that users and items are mapped into a common space and thus they are directly comparable. Extensive experiments on five real-world datasets show significant improvements of DLFM-HSM over the state-of-the-art methods and demonstrate the effectiveness of our model for alleviating the data sparsity problem.