Distance Metric Learning based on Global Cluster Validity Measure by Differential Evolution

Abstract

Distance metric greatly affects the performance of data mining tasks, such as clustering or classification. This paper proposes a distance metric learning based on a global cluster validity measure that evaluates inter- and intra- clusters simultaneously. The proposed method optimizes a distance transform matrix based on Maharanobis distance by utilizing an evolutional algorithm of Differential Evolution (DE). Apart from the most of distance metric learnings, our approach directly improves clustering performance and needs less auxiliary information in principle. In the experiments, we validated the search efficiency of DE, the generalization performance via cross-validation, and also showed how the distance metric learning improves data distribution via visualization by Self-Organizing Map (SOM).