Socially-Aware Self-Supervised Tri-Training for Recommendation

Abstract

Self-supervised learning (SSL), which can automatically generate ground-truth samples from raw data, holds vast potential to improve recommender systems. Most existing SSL-based methods perturb the raw data graph with uniform node/edge dropout to generate new data views and then conduct the self-discrimination based contrastive learning over different views to learn generalizable representations. Under this scheme, only a bijective mapping is built between nodes in two different views, which means that the self-supervision signals from other nodes are being neglected. Due to the widely observed homophily in recommender systems, we argue that the supervisory signals from other nodes are also highly likely to benefit the representation learning for recommendation. To capture these signals, a general socially-aware SSL framework that integrates tri-training is proposed in this paper. Technically, our framework first augments the user data views with the user social information. And then under the regime of tri-training for multi-view encoding, the framework builds three graph encoders (one for recommendation) upon the augmented views and iteratively improves each encoder with self-supervision signals from other users, generated by the other two encoders. Since the tri-training operates on the augmented views of the same data sources for self-supervision signals, we name it self-supervised tri-training. Extensive experiments on multiple real-world datasets consistently validate the effectiveness of the self-supervised tritraining framework for improving recommendation.

Are Graph Augmentations Necessary? Simple Graph Contrastive Learning for Recommendation

“…Following [28], it removes the redundant operations such as the transformation matrices and the nonlinear activation functions. Such a design is proved efficient and effective, and inspires a lot of follow-up CL-based recommendation models like SGL [29] and MHCN [39].…”

“…Inspired by the achievements of CL in other fields, there has been a wave of new research that integrates CL with recommendation [19,29,33,39,41,45]. Zhou et al [45] adopted random masking on attributes and items to create sequence augmentations for sequential model pretraining with mutual information maximization.…”

“…SGL [29] unifies the CL task and the recommendation task under a joint learning framework where the data augmentation is based on the stochastic perturbations on graphs including the node/edge dropout and random walk. SEPT [39] and COTREC [32] further proposes to mine multiple positive samples with semi-supervised learning on the perturbed graph for social/session-based recommendation.…”

“…As the data annotation is not required in CL, it is a natural antidote to the data sparsity issue in recommender systems [22]. An increasing number of very recent studies [29,33,39,41,44,45] have sought to harness CL for improving recommendation performance * Corresponding author. and have demonstrated significant gains.…”

“…Intuitively, we expect that contrasting different graph augmentations can capture the essential information existing in the original user-item interaction, by randomly removing the redundancy and impurity with the edge/node dropout. Unexpectedly, a few latest works [16,39,46] have reported that the dropout rate only has a trivial impact on the performance, and changing it will not cause distinct performance fluctuation. To be more specific, even an extremely sparse graph augmentation (with dropout rate 0.9) in CL can bring decent recommendation performance gains.…”