Social Biases in Wikidata Knowledge Graph

About

In this study, we explored how gender biases encoded in the Wikidata Knowledge Graph propagate to downstream tasks. We considered the link prediction task to classify if a person has a certain occupation. In order to identify occupations which are gender-biased, we utilized fairness metrics like Equalized Odds and Equalized Opportunity. We analyze gender biases across 21 countries to study the variation of biases across different countries. Finally, to study the impact of socio-economic conditions of a country on the occupation-related gender biases, we utilize the Spectral Clustering algorithm. Our findings reveal a striking alignment between the clusters learnt using ML approaches and the established global division of developing (the Global South) and developed countries (the Global North).

Code and Research Paper

Will be released after our paper gets accepted.