Auto-weighted Horizontal Collaboration Fuzzy Clustering

Abstract

Horizontal Collaboration Fuzzy C-Means (HC-FCM) is such a clustering method that does clustering in a set of patterns described in some feature space by considering some external sources of clustering information which are about the same set of patterns but described in different feature spaces. Because of potential security and privacy restrictions, the external information is provided only by some partition matrices obtained by Fuzzy C-Means (FCM). HC-FCM quantifies the influence of these partition matrices by constructing a new objective function which combines these partition matrices with suitable weights. These weights are crucial in HC-FCM, but how to determine them still remains a problem. This paper first puts forward a concept of similarity measure of partition matrices, then based on this similarity measure, proposes some simple and practical methods for automatically determining the weights, and finally gives a new version of HC-FCM, named auto-weighted HC-FCM. With the work of this paper, HC-FCM becomes more practical. Some experiments are carried to show the performance and reveal the validity of our method.