AutoRec++: Incorporating Debias Methods into Autoencoder-based Recommender System

Cheng Liang, Yi He, Teng Huang, Di Wu

Abstract: The deep neural network-based (DNN-based) model has proven powerful in user data behavior representation, efficiently implementing a recommender system (RS). Most prior works focus on developing a sophisticated architecture to better-fit user data. However, user behavior data are commonly collected from multiple scenarios and generated by numerous users, resulting in various biases existing in these data. Unfortunately, prior DNN-based RSs dealing with these biases are fragmented and lack a comprehensive solution. This paper aims to comprehensively handle these biases in user behavior data in preprocessing stage and training state. By incorporating the preprocessing bias (PB) and training bias (TB) into the representative autoencoder-based AutoRec model, we proposed AutoRec++. Experimental results in five commonly used benchmark datasets demonstrate that: 1) the basic model’s preference can boost by the optimal PB and TB combinations, and 2) our proposed AutoRec++ reaches a better prediction accuracy than DNN-based and non-DNN-based state-of-the-art models.