Multi-behavior Self-supervised Learning for Recommendation

Abstract

Modern recommender systems often deal with a variety of user interactions,\ne.g., click, forward, purchase, etc., which requires the underlying recommender\nengines to fully understand and leverage multi-behavior data from users.\nDespite recent efforts towards making use of heterogeneous data, multi-behavior\nrecommendation still faces great challenges. Firstly, sparse target signals and\nnoisy auxiliary interactions remain an issue. Secondly, existing methods\nutilizing self-supervised learning (SSL) to tackle the data sparsity neglect\nthe serious optimization imbalance between the SSL task and the target task.\nHence, we propose a Multi-Behavior Self-Supervised Learning (MBSSL) framework\ntogether with an adaptive optimization method. Specifically, we devise a\nbehavior-aware graph neural network incorporating the self-attention mechanism\nto capture behavior multiplicity and dependencies. To increase the robustness\nto data sparsity under the target behavior and noisy interactions from\nauxiliary behaviors, we propose a novel self-supervised learning paradigm to\nconduct node self-discrimination at both inter-behavior and intra-behavior\nlevels. In addition, we develop a customized optimization strategy through\nhybrid manipulation on gradients to adaptively balance the self-supervised\nlearning task and the main supervised recommendation task. Extensive\nexperiments on five real-world datasets demonstrate the consistent improvements\nobtained by MBSSL over ten state-of-the art (SOTA) baselines. We release our\nmodel implementation at: https://github.com/Scofield666/MBSSL.git.\n