Fair Clustering via Equalized Confidence

Junyi Chai, Xiaoqian Wang

Keywords: fair clustering, Sinkhorn divergence, equalized confidence

TL;DR: fair clustering based on equality of predicted confidence between different demographic groups

Abstract: Fair clustering aims at eliminating effects of sensitive information in clustering assignment. Existing work on fair clustering addresses this problem as a vanilla clustering with constraints that the distribution of protected groups on each cluster should be similar. However, existing criteria for fair clustering does not take into account clustering accuracy, and may restrain the performance of clustering algorithms. To tackle this problem, in this work, we propose a novel metric, equalized confidence, for fair clustering based on the predicted clustering confidence. Instead of enforcing similar distribution of sensitive attributes across different clusters, equalized confidence requires similar predicted confidence across different sensitive groups, bypassing the problem of disparities in statistical features across demographic groups. In light of the new metric, we propose a fair clustering method to learn a fair and good representation for clustering. Compared with conventional methods on fair clustering which try to adjust clustering assignment, our method focuses on learning a fair representation for downstream tasks. Our method proposes to eliminate the disparities of predicted soft labels of samples in different demographic groups using Sinkhorn divergence, as well as to learn clustering-favorable representations for clustering. Experimental results show that our method performs better or comparably than state-of-the-art methods, and that our proposed metric fits better under clustering accuracy.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Social Aspects of Machine Learning (eg, AI safety, fairness, privacy, interpretability, human-AI interaction, ethics)