FairHELP: Fairness-Aware Heterogeneous Information Network Embedding for Link Prediction

Abstract

Heterogeneous information networks (HINs) are ubiquitous in real-world social systems. To effectively learn representations of HINs, Graph Neural Networks (GNNs) have been widely studied as a powerful tool. Nevertheless, there is growing concern that GNNs are prone to make biased predictions in critical decision-making scenarios such as link prediction and social recommendation. Despite recent progress on fair graph learning, few attempts have been made toward promoting fairness in HIN embedding models. In this paper, we study the problem of mitigating link prediction bias in HINs. First, we formalize the definition of fairness in link prediction in HINs, and design fairness measures for the link prediction task. Second, we propose a flexible and model-agnostic debiasing framework named FairHELP for learning fair embeddings in HINs. Third, we conduct extensive experiments on three real-world datasets. The results validate the effectiveness of the proposed fairness measures and the FairHELP framework in achieving fair and accurate link prediction results.