Abstract
Feature selection aims to consistently identify an optimal subset of features that effectively represents the entire dataset or enhances performance in downstream tasks. While deep learning-based approaches have made significant progress in feature selection, they continue to face key challenges, including instability in selected features, limited receptive fields in feature-selection layers due to architectural constraints, and suboptimal utilization of available sample information. To address these limitations, we propose the Robust Fractal Autoencoder (RFAE), an enhanced variant of the Fractal Autoencoder (FAE) designed to improve feature selection stability and adaptability. RFAE introduces three critical advancements: 1) Novel utilization of weight exponentiation to rectify the concern of FAE selecting a reduced number of features than designated. 2) Adoption of a dynamic and tailored strategy to optimize feature selection weights during the training process. 3) Introduction of a optional classification module, facilitating extension to supervised feature selection scenarios. We systematically evaluate RFAE against 14 established feature selection methods. Our experiments span 14 publicly available benchmark datasets, a large-scale GEO gene expression dataset, and a synthetic dataset with known ground-truth features. The results demonstrate that RFAE consistently selects features that achieve lower reconstruction errors while ensuring higher stability across repeated experiments, highlighting its robustness and effectiveness in feature selection tasks.