Regresar

A Comparative Exploration of PCA Variants for Clustering Analysis

Abstract:

In this study, we compare the performance of Principal Component Analysis (PCA), Sparse PCA (SPCA), Robust PCA (RPCA), and Weighted PCA (WPCA) on a high-dimensional dataset of economic indicators from G20 countries. We evaluate their effectiveness in retaining variance and enhancing the performance of K-means clustering. Our comparative analysis employs metrics including effectiveness of variance retention, mean variance of distance sample-centroid, mean distance among centroids, and the rand index for cluster similarity. Our analysis indicates that PCA exhibits a greater effectiveness compared to SPCA but is outperformed by RPCA and significantly by WPCA, which shows the highest variance retention among the four methods. In terms of clustering, SPCA coupled with K-means achieves the best balance between cluster compactness and separation, as indicated by a low mean variance of distance sample-centroid and a relatively high mean distance among centroids. RPCA, while exhibiting extremely compact clusters, demonstrates the least inter-cluster separation. The rand index comparisons reveal that while PCA, SPCA, and WPCA share similar clustering structures, RPCA distinguishes itself by detecting unique patterns, contributing to a broader perspective in the analysis of the high-dimensional datasets. The study provides insightful findings that emphasize the role of appropriate dimensionality reduction method selection in enhancing the effectiveness of unsupervised learning tasks.