Manual versus Data-Driven Specification in Hybrid Discrete Choice Models} - A Benchmark of MNL, RUMBoost, and Penalized Multinomial Logistic Tree Regression (PMLTR)

<title>Abstract</title> <p>Hybrid models that combine discrete choice modeling with tree-based machine learning have emerged as a promising direction for predictive choice analysis. Their adoption within the community has nonetheless been limited, as machine learning methods rarely offer a familiar entry point for analysts trained in classical choice modeling. In addition, their estimated utilities typically lack plain, readable coefficients, so they are hard to interpret and to embed in fields like choice-based optimization. This paper addresses both by proposing Penalized Multinomial Logistic Tree Regression (PMLTR), a hybrid composed of four components, each standard in its discipline: a multinomial logit model, a decision tree to discover breakpoints for binary variables and to identify attribute interactions, a Pearson correlation-based filter for variable pre-selection, and Adaptive Lasso penalization for the final variable selection. First, we compare the PMLTR against the classical MNL model and the dense gradient-boosted RUMBoost architecture. Second, we analyze the impact of manual specifications on the two hybrid approaches. And third, we showcase the possibility of recovering the ground truth of a known utility function. For this, three established datasets from the literature (Swissmetro, London Passenger Mode Choice, and School Trip Dresden) and five synthetic ones are used, spanning a nonlinearity-complexity ladder that demonstrates coefficient interpretability and recovery under known ground truth. All comparisons, predictive and inferential, are backed by respondent-clustered bootstrap tests. Our findings highlight that a careful manual specification, in a MNL or hybrid, can still compete with the purely data-driven alternatives, yet hardly improves upon them. Where prediction is the sole objective, RUMBoost remains the stronger choice, while PMLTR delivers a directly readable utility for interpretation and reuse in applications such as choice-based optimization.</p>