Are Knowledge Graph Embedding Models Biased, or Is it the Data That They Are Trained on?
Wessel Radstok, Melisachew Wudage Chekol, Mirko T. Schäfer · Utrecht University Repository (Utrecht University) · 2021
Recent studies on bias analysis of knowledge graph (KG) embedding models focus primarily on altering the models such that sensitive features are dealt with differently from other features. The underlying implication is that the models cause bias, or that it is their task to solve it. In this paper we argue that the problem is not caused by the models but by the data, and that it is the responsibility of the expert to ensure that the data is representative for the intended goal. To support this claim, we experiment with two different knowledge graphs and show that the bias is not only present in the models, but also in the data. Next, we show that by adding new samples to balance the distribution of facts with regards to specifc sensitive features, we can reduce the bias in the models.