Fair Bayesian Model-Based Clustering

Fairness in clustering is now a practical concern: The paper addresses "group fairness" in clustering—ensuring that sensitive groups (like different demographic categories) are represented proportionally across all clusters rather than concentrated in some clusters and absent from others, reflecting growing demands for trustworthy AI systems.

Existing fair clustering methods have significant limitations: Most current fair clustering approaches are based on K-means and require users to pre-specify both the distance metric and the exact number of clusters in advance, which is often impractical in real-world applications where these parameters are unknown.

The proposed Fair Bayesian Clustering (FBC) method automatically determines cluster numbers: Unlike existing approaches, FBC can infer the optimal number of clusters from data itself, reducing the need for manual parameter tuning and making it more practical for practitioners.

The method works with diverse data types: FBC can be applied to categorical data and other data types beyond numerical data, whereas most existing fair clustering methods are limited to continuous data, expanding its real-world applicability.

Fairness and utility can be balanced competitively: Experiments demonstrate that FBC achieves a competitive trade-off between clustering quality (utility) and fairness, meaning organizations don't necessarily have to sacrifice one for the other when implementing fair clustering systems.