Discrete Choice Modeling using Symbolic Regression with LLM Integration

Abstract

Discrete Choice Modeling (DCM) serves as a cornerstone in understanding decision-making behaviors across transportation, economics, and marketing. While random utility methods offer interpretability, they often suffer from rigid functional form assumptions. Conversely, machine learning approaches maximize predictive power at the cost of interpretability and the risk of overfitting. This research introduces a novel framework integrating Symbolic Regression (SR) with domain knowledge to automate the discovery of utility functions that are both accurate and interpretable. Within the optimization of this framework, we propose a hybrid evolutionary algorithm where an LLM acts as an intelligent operator for population initialization, mutation, and crossover. While the overarching framework remains a Symbolic Regression search, we incorporate the LLM specifically to enhance the generation of candidate solutions. Results on the Swissmetro and Synthetic datasets demonstrate that our SR-DCM implementation is on par with existing machine learning and hybrid models in predictive performance, while significantly reducing overfitting. Furthermore, it outperforms existing interpretable models, with LLM integration serving to further accelerate convergence.