FairCoRe: Fairness-Aware Recommendation Through Counterfactual Representation Learning

Eliminating bias from data representations is crucial to ensure fairness in recommendation. Existing studies primarily focus on weakening the correlation between data representations and sensitive attributes, yet may inadvertently steer the user representations toward another potential bias direction of the target attribute. Furthermore, they often overlook the impact of user preferences on capturing sensitive information, incurring inadequate bias elimination. In this paper, we propose a Fair Counterfactual Representations (FairCoRe) learning framework, which aims to ensure the neutrality of representations among all bias directions. First, we intervene on sensitive attributes to construct a counterfactual scenario. Then, two opposing attribute prediction tasks are respectively performed in ground-truth and counterfactual scenarios to encode sensitive information along different bias directions. Second, we design a bias-aware enhancement learning method that quantifies the respective correlation of user preferences and sensitive attributes to enhance sensitive information encoding. Finally, we introduce two mutual information optimization methods that optimize the representations to capture users’ interests and disentangle sensitive factors. Moreover, we propose an attribute neutralization strategy that refines the learned representations, ensuring sensitive attribute neutrality. Extensive experiments demonstrate that our method achieves the optimal fairness and competitive accuracy compared to state-of-the-art methods.