A deep neural network two-part model and feature importance test for semi-continuous data

This paper addresses the challenge of modeling semi-continuous data where some subjects experience varying degrees of outcomes while others experience none, indicating two distinct data processes. We derive a DNN-based two-part model by adding a bootstrapping procedure along with a filtering algorithm to boost the stability of conventional DNN (sDNN). To improve interpretability, we further derive a feature importance testing procedure to identify important features contributing to the outcome measurements of the two data processes (fsDNN).

Recommended citation: Zou B, Mi X, Wan S, Wu D, Xenakis JG, Hu J, Zou F. (2025). "A deep neural network two-part model and feature importance test for semi-continuous data." Annals of Applied Statistics. 19(2).
Download Paper