Penalized and Weighted K-means for Clustering with Noises and Prior Information Incorporation

Abstract

In this talk a class of loss function extended from K-means criterion is introduced for clustering. Two major extensions are involved: penalization and weighting. The additive penalty term is used to allow a set of noise (scattered) objects without being clustered. Weights are introduced to account for prior information of preferred or prohibited cluster patterns to be found. Their relationships with classification likelihood of Gaussian models are explored. Applications on simulated data as well as real data from microarray and tandem mass spectrometry experiments are evaluated to demonstrate its flexibility and applicability to clustering large complex data.