Towards Fair Representation Learning in Knowledge Graph with Stable Adversarial Debiasing

Yihe Wang, Mohammad Mahdi Khalili, Xiang Zhang

Abstract: With graph-structured tremendous information, Knowledge Graphs (KG) aroused increasing interest in academic research and industrial applications. Recent studies have shown demographic bias, in terms of sensitive attributes (e.g., gender and race), exist in the learned representations of KG entities. Such bias negatively affects specific populations, especially minorities and underrepresented groups, and exacerbates machine learning-based human inequality. Adversarial learning is regarded as an effective way to alleviate bias in the representation learning model by simultaneously training a task-specific predictor and a sensitive attribute-specific discriminator. However, due to the unique challenge caused by topological structure and the comprehensive relationship between knowledge entities, adversarial learning-based debiasing is rarely studied in representation learning in knowledge graphs. In this paper, we propose a framework to learn unbiased representations for nodes and edges in knowledge graph mining. Specifically, we integrate a simple-but-effective normalization technique with Graph Neural Networks (GNNs) to constrain the weights updating process. Moreover, as a work-in-progress paper, we also find that the introduced weights normalization technique can mitigate the pitfalls of instability in adversarial debasing towards fair-and-stable machine learning. We evaluate the proposed framework on a benchmarking graph with multiple edge types and node types. The experimental results show that our model achieves comparable or better gender fairness over three competitive baselines on Equality of Odds. Importantly, our superiority in the fair model does not scarify the performance in the knowledge graph task (i.e., multi-class edge classification).