Geometric divergence based fuzzy clustering with strong resilience to noise features

In this article we consider the problem of fuzzy partitional clustering using a separable multi-dimensional version of the geometric distance which includes f-divergences as special cases. We propose an iterative relocation algorithm for the Fuzzy C Means (FCM) clustering that is guaranteed to converge to local minima. We also demonstrate, through theoretical analysis, that the FCM clustering with the proposed divergence based similarity measure, is more robust towards the perturbation of noise features than the standard FCM with Euclidean distance based similarity measure. In addition, we show that FCM with the suggested geometric divergence measure has better or comparable clustering performance to that of FCM with squared Euclidean distance on real world and synthetic datasets (even in absence of the noise features).