Principal Components Clustering through a Variance-Defined Metric

Abstract

This work aims at proposing a clustering procedure through a new metric, a weighted Euclidean distance, in which the weights are the ratio of corresponding eigenvalues and the largest eigenvalue found after a Principal Components Analysis. In order to illustrate the method, the procedure was carried out on twenty-one newborn EEG segments, classified as TA (Tracé Alternant) or HVS (High Voltage Slow) patterns. The observed clustering structure was assessed by the cophenetic and agglomerative coefficients. Results showed that, despite its unlikely existence, a clustering structure was suggested by the traditional approach. This structure, however, was not confirmed by the proposed method.