Temporal Contrastive Pre-Training for Sequential Recommendation

Abstract

Recently, pre-training based approaches are proposed to leverage self-supervised signals for improving the performance of sequential recommendation. However, most of existing pre-training recommender systems simply model the historical behavior of a user as a sequence, while lack of sufficient consideration on temporal interaction patterns that are useful for modeling user behavior. In order to better model temporal characteristics of user behavior sequences, we propose a Temporal Contrastive Pre-training method for Sequential Recommendation (TCPSRec for short). Based on the temporal intervals, we consider dividing the interaction sequence into more coherent subsequences, and design temporal pre-training objectives accordingly. Specifically, TCPSRec models two important temporal properties of user behavior, i.e., invariance and periodicity. For invariance, we consider both global invariance and local invariance to capture the long-term preference and short-term intention, respectively. For periodicity, TCPSRec models coarse-grained periodicity and fine-grained periodicity at the subsequence level, which is more stable than modeling periodicity at the item level. By integrating the above strategies, we develop a unified contrastive learning framework with four specially designed pre-training objectives for fusing temporal information into sequential representations. We conduct extensive experiments on six real-world datasets, and the results demonstrate the effectiveness and generalization of our proposed method.