Debiased Contrastive Learning for Sequential Recommendation

Abstract

Current sequential recommender systems are proposed to tackle the dynamic\nuser preference learning with various neural techniques, such as Transformer\nand Graph Neural Networks (GNNs). However, inference from the highly sparse\nuser behavior data may hinder the representation ability of sequential pattern\nencoding. To address the label shortage issue, contrastive learning (CL)\nmethods are proposed recently to perform data augmentation in two fashions: (i)\nrandomly corrupting the sequence data (e.g. stochastic masking, reordering);\n(ii) aligning representations across pre-defined contrastive views. Although\neffective, we argue that current CL-based methods have limitations in\naddressing popularity bias and disentangling of user conformity and real\ninterest. In this paper, we propose a new Debiased Contrastive learning\nparadigm for Recommendation (DCRec) that unifies sequential pattern encoding\nwith global collaborative relation modeling through adaptive conformity-aware\naugmentation. This solution is designed to tackle the popularity bias issue in\nrecommendation systems. Our debiased contrastive learning framework effectively\ncaptures both the patterns of item transitions within sequences and the\ndependencies between users across sequences. Our experiments on various\nreal-world datasets have demonstrated that DCRec significantly outperforms\nstate-of-the-art baselines, indicating its efficacy for recommendation. To\nfacilitate reproducibility of our results, we make our implementation of DCRec\npublicly available at: https://github.com/HKUDS/DCRec.\n