A Hybrid Approach for Neural Collaborative Filtering

Recommender systems enable better personalization for e-commerce and online entertainment services and thus gain significant attention from researchers. Alongside the traditional matrix factorization approach, neural networks have been recently a promising trend for collaborative filtering-based systems thanks to considerable improvements in the quality of the recommendations. This paper introduces a hybrid collaborative filtering framework that applies both approaches in a parallel manner to learn knowledge from implicit feedback data. Embedding vectors representing the information of users and items are first mapped from data. Matrix factorization is generalized by the element-wise product of these embeddings, while the neural network takes as input a 2-D interaction map formed from the stacking of two vectors. The framework fuses the element outputs by concatenation to produce an accurate estimation of the correlation between users and items. The proposed method outperformed several baselines in the experiments on standard datasets, including MovieLens, Yelp, and Pinterest. This advantage suggests more considerations on the integration of deep learning to collaborative filtering for effective recommender systems.