A Study on Improvement of Clustering Performance for Hierarchical Topological Clustering based on Adaptive Resonance Theory

Abstract

Hierarchical clustering can extract knowledge with various information granularities from data, and various methods have been studied. One of them is a hierarchical model of CIM-based ART with Edge and Age (CAEA) based on the Adaptive Resonance Theory (HCAEA). HCAEA automatically calculates a similarity threshold from the distribution of data points. However, it has the problem of excessive node generation. In this paper, the number of initial nodes generated in the second and subsequent layers is changed, and a new node generation criterion is introduced to suppress excessive node generation. Furthermore, we aim to improve the clustering performance by reusing the data discarded with the deletion of nodes in the training process of HCAEA. Numerical experiments on real-world data confirmed that the proposed method improves clustering performance.