Title: K-bMOM: A robust K-means-type procedure with application to color quantization
Authors: Edouard Genetay - CREST-ENSAI (France) [presenting]
Adrien Saumard - Crest-Ensai (France)
Camille Saumard - artfact lumenAI (France)
Abstract: Classical clustering methods, such as K-means, suffer from a lack of robustness with respect to outliers. We propose a robust version of K-means named K-bMOM, using bootstrap and median-of-means statistics, a strategy that has been recently put to emphasis for efficient, robust machine learning. The algorithm is iterative, in a Lloyd-type fashion. The performances of K-bMOM are theoretically and empirically shown. First, we give a theoretical majoration of the risk excess. Secondly, simulations show that K-bMOM converges rapidly along with the iteration steps, that it clearly outperforms K-means on corrupted or heavy-tailed data and that it is competitive with other robust approaches, such as K-median for instance. K-bMOM also provides interesting outcomes such as a robust and efficient initialisation procedure and outlier detection.