RESETBERT4Rec: A Pre-training Model Integrating Time And User Historical Behavior for Sequential Recommendation

Abstract

Sequential recommendation methods are very important in modern recommender systems because they can well capture users' dynamic interests from their interaction history, and make accurate recommendations for users, thereby helping enterprises succeed in business. However, despite the great success of existing sequential recommendation-based methods, they focus too much on item-level modeling of users' click history and lack information about the user's entire click history (such as click order, click time, etc.). To tackle this problem, inspired by recent advances in pre-training techniques in the field of natural language processing, we build a new pre-training task based on the original BERT pre-training framework and incorporate temporal information. Specifically, we propose a new model called the RE arrange S equence prE -training and T ime embedding model via BERT for sequential R ecommendation (RESETBERT4Rec ) \footnoteThis work was completed during JD internship., it further captures the information of the user's whole click history by adding a rearrange sequence prediction task to the original BERT pre-training framework, while it integrates different views of time information. Comprehensive experiments on two public datasets as well as one e-commerce dataset demonstrate that RESETBERT4Rec achieves state-of-the-art performance over existing baselines.