Incorporating Domain Knowledge in Deep Neural Networks for Mode Choice Analysis

Abstract

Discrete choice models (DCM) are widely used in travel demand analysis to understand and predict choice behaviors. However, a priori specification of the utility functions is required for model estimation, leading to subjectivity and potential inaccuracies. Machine learning (ML) approaches have emerged as a promising solution but lack interpretability and may not capture expected relationships. This study proposes a framework that supports the development of interpretable models that incorporate domain knowledge and prior beliefs. The framework includes pseudo data samples and a loss function to measure relationship fulfillment. This approach combines the flexibility of ML structures with econometrics and interpretable behavioral analysis, improving model interpretability. The proposed framework's potential is demonstrated through a case study, providing a promising avenue for the advancement of datadriven approaches in DCM.