Unsupervised Feature Selection and Deep Subspace Clustering for Exploratory High-Dimensional Cluster Analysis

Abstract

With the advancement of information technology, data volume is rapidly increasing, po- sing significant challenges for storage and processing. This growth occurs both in the

number of samples and in the number of features, making initial exploratory small data

analysis crucial to reducing computational demands and improving data quality for ma- chine learning (ML) training. However, simply reducing the number of samples can in- tensify the “curse of dimensionality,” complicating analysis when a small dataset contains

many features. Dimensionality reduction techniques are therefore essential for enabling more efficient and interpretable analyses. Unlike methods such as PCA, which transform

the original data, unsupervised feature selection techniques identify the most relevant va- riables without requiring labels, enhancing the interpretability of natural data patterns.

However, patterns may emerge only within specific feature subsets, known as subspaces. In some cases, the original features may not be sufficient, requiring the generation of new

ones to identify these subspaces. This research explores two strategies for handling high- dimensional data with few samples: (i) a novel unsupervised feature selection method and

(ii) a clustering approach based on subspaces. Experiments on real and synthetic datasets showed that the proposed methods outperform state-of-the-art approaches, as evidenced by clustering evaluation metrics and statistical tests.