Abstract: With the advancement of big data technologies, high-dimensional datasets have seen widespread application, often accompanied by redundancy and noise that lead to overfitting and high computational costs. Feature selection addresses these issues by identifying informative subsets to reduce dimensionality and enhance model interpretability. While deep neural networks excel in feature extraction, they are sensitive to noise and outliers. To overcome this, we propose a novel feature selection approach based on an Intuitionistic Fuzzy Deep Autoencoder (IFDAE). By combining fuzzy theory with a deep autoencoder, IFDAE manages uncertainty by computing intuitionistic fuzzy weights from membership and non-membership information, reducing noise effects. It then learns nonlinear latent features and transfers pre-trained weights to the target network. A structured sparse regularization applied to the input-to-hidden connections yields final feature importance scores. Experiments on 12 public datasets and a real-world schizophrenia dataset confirm IFDAE’s strong reconstruction and classification capabilities.