Gaussian Mixture Models Meet Econometric Models

Since McFadden linked his logit formulation with random utility theory in the 1970s, discrete choice modelers have been mostly relying on the multinomial logit model (MNL) and its variants due to the connection provided to economic theory. In the last decades, the use of machine learning (ML) methods has grown rapidly in many fields such as mobile/web apps, computer vision, natural language processing, and robotics, to name a few. As for transportation, machine learning has been applied to the area of traffic control, traffic forecasting, incident detection, and prediction of transportation mode from raw GPS/mobile phone data. However, discrete choice modelers, mainly econometricians, still count on econometric models instead of machine learning. This may be due to the missing link with economic theory and the lack of simple interpretability that econometric models usually provide. This research tries to combine machine learning with traditional discrete choice models (DCM) in a hybrid framework that makes use of the advantages of both fields while maintaining the economic interpretability of the choice model. In particular, Gaussian Mixture Models (GMM), an unsupervised machine learning algorithm, is added to the traditional Latent Class Choice Model (LCCM) as an alternative to the class membership model that classifies/clusters people into homogenous groups. Results show that the proposed hybrid approach provides similar goodness-of-fit measures compared to LCCM with slightly better prediction accuracy. In addition, the new approach is able to identify more latent classes than LCCM.