Fairness Through Controlled (Un)Awareness in Node Embeddings

Abstract

Graph representation learning is crucial for applying many machine learning (ML) models to complex real-world graphs, like social networks. Ensuring ‘fair’ representations is essential, due to the the societal implications of such ML systems that also often use sensitive personal data. This work evaluates the CrossWalk algorithm, which was designed for fair representations, on multiple social-network datasets. We analyze representation quality for both sensitive and non-sensitive attributes with respect to multiple quality metrics. Our study shows CrossWalk’s adaptability to different fairness paradigms and demonstrates that benefits of these fairness interventions are more pronounced for underrepresented groups.