FairMB: A Multi-behavior Group Fairness Model for Recommendation with Representation Augmentation
Abstract
Ensuring fairness in modern recommender systems is an ongoing imperative, yet current methodologies fail to effectively capture the complex interplay between long-tail distributions and biases. Specifically, sparse interactions cause tail user nodes to undergo semantic homogenization, progressively entangling their preferences with sensitive attributes. Consequently, such entanglement severely degrades recommendation utility and fairness for tail users. In this paper, we propose FairMB, a multi-behavior fairness paradigm designed to mitigate the tail-node degradation by leveraging rich collaborative signals from auxiliary behaviors. Specifically, a tail-aware aggregation mechanism exploits additional auxiliary signals to provide tail-node compensation. Furthermore, to prevent additional biases from auxiliary behaviors, a dual-guided debiasing augmentation network is employed. Driven by downstream tasks and projection?based purified residual embeddings, it constructs fair augmented targets for contrastive alignment, guiding the original representations to disentangle biases from preferences and ultimately achieving an optimal utility-fairness trade-off. Comprehensive evaluations across diverse recommendation benchmarks confirm that FairMB effectively achieves accuracy-fairness trade-off, establishing state-of-the-art standards for fair representation learning in multi-behavior environments.