A New Multi-Feature Recommendation Model Based on Recurrent Neural Network

Abstract

With the problems of handling single-feature and overlooking user preferences in the recommendation algorithms, this paper proposes a Recurrent Neural NetWork-based Multi-feature Hybrid Recommendation Model (RN-MHRM). Firstly, features are extracted from user-item interaction data using the Latent Factor Model (LFM), and an improved Recurrent Neural NetWork (RNN) is used to replace the linear inner product of LFM vectors With non-linearity, Which aims at learning richer features that capture user's short-term interests. Secondly, to avoid single-feature, item information is introduced and the BERT model is used for extracting multi-features. Thirdly, both short-term and long-term interests are considered, and the user's long-term interests are trained by a FeedforWard Neural NetWork (FNN), Which greatly improves the recommendation performance. Experiments designed on multiple real datasets have shown that RN-MHRM effectively improves recommendation performance compared to the baseline model.