Swanand Ravindra Kadhe, Farhan Ahmed, et al.
ICML 2024
This paper considers centroid-based clustering under ℓ1 distance in which both data points and cluster centers are subject to a sum constraint on their components. A closed-form solution is derived for the cluster center optimization problem, enabling an interpretation as a sample quantile of the cluster. An adaptive sampling initialization step is also adopted to provide a guarantee on expected clustering cost as well as empirical improvements. Experiments on synthetic data indicate that the advantages of the proposed algorithms increase as clusters become more concentrated and as the dimension increases. An application to clustering employee job role profiles highlights the utility of ℓ1 distance in promoting sparse, interpretable cluster centers.
Swanand Ravindra Kadhe, Farhan Ahmed, et al.
ICML 2024
Changsheng Wang, Yihua Zhang, et al.
ICML 2025
Zirui Yan, Dennis Wei, et al.
AISTATS 2024
Oznur Alkan, Dennis Wei, et al.
MLSys 2022