Factorization Meets Neural Networks: A Scalable and Efficient Recommender for Solving the New User Problem

Abstract

Predicting new user behavior has always been a challenging issue in intelligent recommender systems. This challenge is mainly due to the extreme asymmetry of information between new users and old users. Existing factorization models can efficiently process and map asymmetric information, but they are not good at mining deep relationships between contexts when compressing high-dimensional data. In contrast, neural network methods can deeply exploit the relationship between contexts; however, their training cost is much larger than factorization approaches. Therefore, this paper proposes a scalable and efficient recommender to solve the new user problem by filling the gap between factorization models and neural networks. The scalable part is a neural network that can jointly encode, compress, and fuse various types of contexts. The efficient part is a factorization model based on a correlation constraint mechanism and a projection strategy, which enables an asymmetric mapping of information between old and new users. The entire recommender fuses the two parts so that the factorization model and the neural network can complement each other. The experimental results show that our approach can achieve a good balance between performance and training efficiency compared to state-of-the-art methods.