Instance-Level Neural Feature Selection Based on Disentanglement Enhancement

Feature selection aims to select the optimal subset from the original features to reduce data dimensionality, eliminate irrelevant features, and improve the interpretability and performance of the model. In recent years, feature selection methods using neural networks have received much attention, but they often lead to suboptimal results due to feature entanglement problems. To this end, we propose a novel instance-level neural feature selection method based on disentanglement enhancement, including a feature selection network, a label prediction network, and a disentanglement bootstrap network. The feature selection network divides the features into strongly and weakly relevant ones, the label prediction network uses the strongly relevant features for supervised learning, and the disentanglement bootstrap network promotes feature disentanglement by minimizing the mutual information between strong and weak features. The designed gradient-based solution algorithm effectively solves the joint optimization problem. Experimental results validate the effectiveness of the method on tabular and image datasets.