Differentiable gated autoencoders for unsupervised feature selection

Unsupervised feature selection (UFS) aims to identify a subset of the most informative features from high- dimensional data without labels. However, most existing UFS methods cannot adequately capture the intricate nonlinear relationships present in high-dimensional data, limiting their ability to obtain meaningful features. In this letter, we propose a novel UFS model that leverages the autoencoder with a differentiable gating function to address this limitation. Our model optimizes a bi-level loss function to automatically select a subset of informative and stable features by capturing the nonlinear interactions between features. Furthermore, we leverage the continuous relaxation of the Bernoulli distribution to parameterize stochastic gates, thereby enabling the learning of a relaxed model through a low-variance gradient estimator. The proposed model does not necessitate prior knowledge or domain-specific tuning, making it a versatile and widely applicable method. Extensive experiments on 13 benchmark datasets demonstrate the effectiveness and superiority of the proposed method.