Fair Neighbor Embedding

We consider fairness in dimensionality reduction (DR). Nonlinear DR yields low dimensional representations that let users visualize and explore high-dimensional data. However, traditional DR may yield biased visualizations overemphasizing relationships of societal phenomena to sensitive attributes or protected groups. We introduce a framework of fair neighbor embedding, the Fair Neighbor Retrieval Visualizer, formulating fair nonlinear DR as an information retrieval task with performance and fairness quantified by information retrieval criteria. The method optimizes low-dimensional embeddings that preserve high-dimensional data neighborhoods without biased association of such neighborhoods to protected groups. In experiments the method yields fair visualizations outperforming previous methods.