A1208
Title: Incorporating group fairness in a variable-weighted adjacency construction for spectral clustering
Authors: Jesse Ghashti - University of British Columbia (Canada) [presenting]
John Thompson - The University of British Columbia (Canada)
Abstract: Constraining clustering algorithms or applying post-hoc modifications to enforce fairness often pushes the objective function away from local or global optima. A nonparametric method is proposed for group-fair bandwidth selection, motivated by kernel density estimation, where estimated bandwidths act as variable-specific scaling factors in a pairwise weighted distance. These distances are then converted to similarities with a kernel function to form the adjacency matrix for both hard and soft spectral clustering. Applications show that the proposed method produces clustering results that are more group-fair than standard spectral clustering, while yielding smaller deviations from the optimum than other fairness-constrained variants. Finally, it is shown that soft clustering provides probabilistic memberships with a more natural interpretation than hard assignments, and extensions to mixed variable types and alternative fairness definitions are discussed.
