Neural Gaussian Mirror for Controlled Feature Selection in Neural Networks

TL;DR

This paper introduces neural Gaussian mirrors (NGMs), a novel method that creates mirrored features through structured perturbations to evaluate feature importance in neural networks, enabling controlled feature selection.

Contribution

The paper proposes a new neural network architecture with mirrored features and mirror statistics for controlled feature selection, addressing limitations in existing DNN interpretability.

Findings

Controls feature selection error rate effectively

Maintains high selection power with correlated features

Validated on simulated and real datasets

Abstract

Deep neural networks (DNNs) have become increasingly popular and achieved outstanding performance in predictive tasks. However, the DNN framework itself cannot inform the user which features are more or less relevant for making the prediction, which limits its applicability in many scientific fields. We introduce neural Gaussian mirrors (NGMs), in which mirrored features are created, via a structured perturbation based on a kernel-based conditional dependence measure, to help evaluate feature importance. We design two modifications of the DNN architecture for incorporating mirrored features and providing mirror statistics to measure feature importance. As shown in simulated and real data examples, the proposed method controls the feature selection error rate at a predefined level and maintains a high selection power even with the presence of highly correlated features.