Clustering Optimization via Centroid Neural Network Ensemble

A cluster analysis approach based on an ensemble of unsupervised competitive neural networks for clustering optimization is proposed in this paper. The proposed method utilizes centroid neural network (CentNN) as its backbone, which has been recognized as an effective data clustering algorithm that can yield more satisfactory clustering error rates compared to conventional methods. In an attempt to further enhance the clustering results, this study investigates a centroid neural network ensemble (CentNN-E). By leveraging the strengths of multiple CentNN models, the proposed CentNN-E scheme aims to mitigate individual biases and achieve more robust and reliable clustering results. For one specific cluster, the prototypes (centroid candidates) produced by multiple CentNN models (with different initializations) are grouped by using Euclidean distance and are convexly combined using error-based weighting on a per-cluster basis. The resulting prototypes from the ensemble are subsequently utilized as initial prototypes for a final execution. Experimental results on various synthetic test data sets demonstrate that the proposed CentNN-E approach can yield superior results and surpass a single CentNN algorithm in terms of clustering error criterion.