Principal Component Analysis Based on Group Sparse Learning

Abstract Dimensionality reduction is a fundamental technique used across various domains such as pattern recognition, machine learning, and bioinformatics, aimed at simplifying data without losing essential information. However, traditional methods often utilize all the features from the original dataset, which can include irrelevant, redundant, or noisy features, thereby affecting the overall model performance. To address these challenges, we propose a novel approach called Group Sparse Principal Component Analysis (GSPCA). This method introduces sparsity constraints on feature groups, enabling selective dimensionality reduction and reducing the impact of irrelevant features. Moreover, we perform both theoretical analysis and extensive experiments, which demonstrate that GSPCA outperforms conventional techniques, offering significant improvements in both computational efficiency and model accuracy. Our method, therefore, presents a robust solution for scenarios requiring efficient handling of high-dimensional data.