Fig 4. Proposed approach to couple RF and count data models (Poisson: POI, negative binomial: NB, zero-inflated Poisson: ZIP, and zero-inflated negative binomial: ZINB).

First, the ensemble of decision trees is used to segment the samples into groups with similar characteristics. These trees are shallow trees, so that each of the leaf nodes contains a few hundred of samples. Second, we plug the selected count data models to each of the leaf nodes in the ensemble. The predictions yielded by the count data models are subsequently averaged to obtain the final prediction for each RF and count data model combination.