Two-Way Sequence Modeling for Context-Aware Recommender Systems with Multiple Interactive Bidirectional Gated Recurrent Unit

Abstract

For modeling the user behavior in recommender systems, the task of combining the contexts of interactions corresponds to the sequential item history has inevitable role in improving the quality of recommendations. The resort of existing recommendation models is the left-to-right autoregressive training approach. While training a certain model at a specific time step, both future (right) context/data along with the past (left) is always available in the given training set sequences. It is intuitive that the current behavior of the user has certain connections with their future actions too. Future behaviors of users can boost the quality of recommendations. In this paper, two-way sequence modeling technique is proposed for concatenating both left-to-right (past) and right-to-left (future) dependencies in a user interaction sequence. Inspired from the text modeling techniques, a Multiple interactive Bidirectional Gated Recurrent Unit (MiBiGRU) architecture is proposed to model the two-way dependencies in recommender systems. Modeling future contexts along with past contexts is an auspicious way for attaining better recommendation accuracy.