Predictive weighting for cluster ensembles

Abstract

An ensemble of regression models predicts by taking a weighted average of the predictions made by individual models. Calculating the weights such that they reflect the accuracy of individual models (post processing the ensemble) has been shown to increase the ensemble's accuracy. However, post processing cluster ensembles has not received as much attention because of the inherent difficulty in assessing the accuracy of an individual cluster model. By enforcing the notion that clusters must be predictable, this paper suggests a means of implicitly post processing cluster ensembles by drawing analogies with regression post processing techniques. The product of the post processing procedure is an intelligently weighted co-occurrence matrix. A new technique, similarity-based k-means (SBK), is developed to split this matrix into clusters. The results using three real life datasets underpinned by chemical and biological phenomena show that splitting an intelligently weighted co-occurrence matrix gives accuracy that approaches supervised classification methods.