Unifying paragraph embeddings and neural collaborative filtering for hybrid recommendation

Abstract

Abstract Collaborative filtering is one of widely used recommendation techniques. Despite the effectiveness of matrix factorization for collaborative filtering; however, the inner product operator, combining the multiplication of latent features linearly, may not be sufficient to capture the complex structure of user interaction ratings. On the other hand, we argue that there is a great deviation between user ratings and their real interest preference. In this paper, we propose a novel hybrid recommendation algorithm. It adopts neural networks to exploit user–item ratings for collaborative filtering, which is endowed a high level of non-linearity for capturing the complex structure of user interaction ratings. At the same time, it exploits item embeddings to capture the content feature for auxiliary information, which solves the cold start problem to some extent. In particular, we introduce paragraph embeddings to represent user reviews and item descriptions, and design two neural networks to capture the sentiment of user reviews and the content feature of items, respectively. And then, we treat these embeddings as attention weights of users and items, and unify them with user–item ratings to model the hybrid recommendation system. Extensive experiments on Amazon product dataset demonstrates that our algorithm performs better on rating prediction than other state-of-the-art algorithms.