A Growing Hierarchical Clustering Algorithm via Parameter-free Adaptive Resonance Theory

Kazuki Tashiro, Naoki Masuyama, Yusuke Nojima

Abstract: Generally, clustering algorithms based on Adaptive Resonance Theory (ART) require the setting of data-dependent parameters such as similarity thresholds. Previous studies have shown that Correntropy-Induced Metric (CIM)-based ART+ (CA+), which automatically determines similarity thresholds based on the diversity of training data, exhibits superior clustering performance. CA+ is a non-hierarchical clustering algorithm that autonomously and adaptively generates nodes by the training data. This paper introduces a hierarchical structure to CA+ to enhance clustering performance. We discuss the clustering performance and clustering characteristics of the proposed algorithm through comparative experiments with other algorithms using real-world datasets.