A heuristic approach to improve the robustness of a railway timetable in a bottleneck area

Abstract

Deep Neural Networks (DNNs) are accurate and powerful tools for modeling travel decisions. Nonetheless, the black-box characteristic of DNNs has decreased their potential implication in discrete choice modeling. In this study, we investigate the potentials of cutting-edge post-hoc interpretation tools in providing behavioral insight into DNN architectures. We evaluate the relationship between the output probabilities and input features using the Shapely Additive explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Using SwissMetro dataset, we demonstrate that the outputs of SHAP and LIME are consistent with theory when the architecture of DNN is designed based on the Random Utility Maximization (RUM) theory. However, for a fully connected DNN architecture, SHAP and LIME do not provide behaviorally interpretable outputs. Additionally, the prediction accuracy shows the DNN model based on RUM avoids overfitting.