Travel behavior modeling is essential for transportation demand analysis and policy-making, yet traditional discrete choice models often struggle with real-world data complexities, such as heavy-tailed distributions and strong feature correlations. This study proposes a novel end-to-end neural network framework integrated with advanced statistical techniques to effectively address these issues. Specifically, a ParetoTail transformation is employed to normalize heavy-tailed travel attributes, such as travel time and cost, reducing the undue influence of extreme values. To explicitly capture complex dependencies among features, a Gaussian copula approach is integrated, improving the robustness of the model against traditional independence assumptions. Furthermore, a gating mechanism is introduced to dynamically balance the contributions of continuous and discrete features, incorporating random noise to account for preference heterogeneity across individual travelers. Using the Swissmetro dataset, extensive empirical analyzes demonstrate that the proposed model significantly outperforms the baseline models (MNL, MXL, L-MNL, E-MNL, EL-MNL) in terms of prediction accuracy, F1 score, and AUC values, highlighting its superior ability to handle long-tailed distributions and feature dependencies. Additional ablation studies underscore the essential roles of the ParetoTail, Gaussian copula, and gating components. In general, this integrated framework provides a flexible and robust approach to modeling travel behavior in complex real-world scenarios.
Swissmetro Dataset （Processed completed）
- swissmetro.dat (3 classes)
- BIOGEME official website.