Adversarial Learning for Debiasing Knowledge Graph Embeddings

Abstract

Knowledge Graphs (KG) are gaining increasing attention in both academia and industry. Despite their diverse benefits, recent research have identified social and cultural biases embedded in the representations learned from KGs. Such biases can have detrimental consequences on different population and minority groups as applications of KG begin to intersect and interact with social spheres. This paper describes our work-in-progress which aims at identifying and mitigating such biases in Knowledge Graph (KG) embeddings. As a first step, we explore popularity bias -the relationship between node popularity and link prediction accuracy. In case of node2vec graph embeddings, we find that prediction accuracy of the embedding is negatively correlated with the degree of the node. However, in case of knowledge-graph embeddings (KGE), we observe an opposite trend. As a second step, we explore gender bias in KGE, and a careful examination of popular KGE algorithms suggest that sensitive attribute like the gender of a person can be predicted from the embedding. This implies that such biases in popular KGs is captured by the structural properties of the embedding. As a preliminary solution to debiasing KGs, we introduce a novel framework to filter out the sensitive attribute information from the KG embeddings, which we call FAN (Filtering Adversarial Network). We also suggest the applicability of FAN for debiasing other network embeddings which could be explored in future work.

Fairness in Recommendation: A Survey

“…Learning. Another typical technique to promote fairness is to take advantage of adversary learning [10,13,17,25,59,62,64,176,197]. The basic idea of mitigating unfairness through adversary learning is to learn fair representations through a min-max game between the main task predictor and an adversarial classifier.…”

Deep Automation Bias: How to Tackle a Wicked Problem of AI?

“…Many scholars from various disciplines have been pointing out that algorithms and particularly ML are biased, unfair and lack in transparency and accountability ( [1][2][3][4][5][6][7][8][9][10][11][12]). To deal with these serious issues, there is important research proceeding in the AI/ML community, particularly in computer science, data science and related disciplines aiming at developing approaches for debiasing and improving FAT ( [3,4,[12][13][14][15]). Correspondingly, there is a growing research community dealing with these issues.…”

“…While it is obviously relevant to tackle these issues of data quality and technical bias, there is also a need for approaches to address the societal and ethical issues of AI. Several scholars thus argue that bias is not merely a technical issue ( [1,[3][4][5][6][11][12][13][14][15][16][17]). The complexity of the problem is already observable in the various different, and partially contradictory notions and definitions of fairness ( [5,11]).…”

Implications of Topological Imbalance for Representation Learning on Biomedical Knowledge Graphs

“…The issue of biased training data has become an area of great interest within the KG field. Typically, the bias of concern is that of attributes associated with the entities in the graph, for example gender, age or race [47,3,8]. However, Recent work has shown how popularity bias is present in three frequently used non-biomedical KGs: FB15K, WN18 and YAGO3-10 [27].…”

“…The issue of non-uniform graph connectivity (typically in homogenous graphs) has begun to be studied in parallel by the field of Graph Neural Networks (GNN), where researchers have shown that models learn low-quality representations, thus making more incorrect predictions, for low-degree vertices [26,25,44]. This has also been explored in the context of homogenous graph representation learning [3] and for random walks [23,36].…”