Contrastive learning with adversarial masking for sequential recommendation

Abstract

Sequential recommendation is of paramount importance for predicting user preferences based on their historical interactions. Recent studies have leveraged contrastive learning as an auxiliary task to enhance sequence representations, with the goal of improving recommendation accuracy. However, an important challenge arises: random item masking, a key component of contrastive learning, while promoting robust representations through intricate semantic inference, may inadvertently distort the original sequence semantics to some extent. In contrast, methods that prioritize the preservation of sequence semantics tend to neglect the essential masking mechanism for robust representation learning. To address this issue, we propose a model called Contrastive Learning with Adversarial Masking (CLAM) for sequential recommendation. CLAM consists of three core components: an inference module, an occlusion module, and a multi-task learning paradigm. During training, the occlusion module is optimized to perturb the inference module in both recommendation generation and contrastive learning tasks by adaptively generating item embedding masks. This adversarial training framework enables CLAM to balance sequential pattern preservation with the acquisition of robust representations in the inference module for recommendation tasks. Our extensive experiments on four benchmark datasets demonstrate the effectiveness of CLAM. It achieves significant improvements in sequential recommendation accuracy and robustness against noisy interactions.