Robust clustering with Gaussian estimator

Abstract

To overcome the drawback of traditional FCM algorithm, i.e., the sensitivity to noise and outliers, a simple but efficient M-estimator, Gaussian estimator, is introduced into clustering analysis. Based on the relationship between robust statistics and clustering analysis, a Robust Gaussian Clustering (RGC) algorithm is presented, which can be viewed as a collection of C-independent Gaussian estimators to achieve robust estimation of the desired C cluster centers. Theoretic study and simulation results show that the RGC algorithm has a clear mathematical meaning and reasonable physical interpretation. It can also obtain efficient and robust estimation of the prototype parameters even when the data set is contaminated by heavy noise. Both the clustering results on a noisy data set and the classification performance on a real data set indicate that this novel algorithm outperforms the traditional FCM algorithm.