---

In the broad landscape of modern data analysis, dealing with count data affected by excessive zeros represents a fundamental analytical challenge. Classical models such as Poisson regression exhibit notable weaknesses in this context, as they cannot distinguish between structural zeros (arising from deterministic processes) and random zeros (arising from stochastic processes). Although zero-inflated Poisson regression models have taken a significant step toward addressing this issue, their performance is seriously limited in the era of high-dimensional and large-scale data.
This research, looking beyond traditional frameworks, introduces an innovative deep neural network–based framework designed to enhance and transform zero-inflated Poisson regression models. By leveraging the remarkable capacity of deep learning to extract nonlinear and complex features, this hybrid approach can model the dual data-generating processes, the zero-generation process, and the count-generation process with unprecedented precision.