Incorporating graph neural network into route choice model

Route choice models are one of the most important foundations for transportation research. Traditionally, theory-based models have been utilized for their great interpretability, such as logit models and Recursive logit models. More recently, machine learning approaches have gained attentions for their better prediction accuracy. In this study, we propose novel hybrid models that integrate the Recursive logit model with Graph Neural Networks (GNNs) to enhance both predictive performance and model interpretability. To the authors’ knowldedge, GNNs have not been utilized for route choice modeling, despite their proven effectiveness in capturing road network features and their widespread use in other transportation research areas. We mathematically show that our use of GNN is not only beneficial for enhancing the prediction performance, but also relaxing the Independence of Irrelevant Alternatives property without relying on strong assumptions. This is due to the fact that a specific type of GNN can efficiently capture multiple cross-effect patterns on networks from data. By applying the proposed models to one-day travel trajectory data in Tokyo, we confirmed their higher prediction accuracy compared to the existing models.

💡 Research Summary

This paper introduces a novel hybrid route‑choice model that integrates the Recursive Logit (RL) framework with Graph Neural Networks (GNNs), named ResDGCN‑RL. Traditional theory‑based models such as the Multinomial Logit (MNL) and its extensions (C‑Logit, Path‑Size Logit, Mother Logit, GEV‑based models) offer high interpretability but suffer from two major drawbacks: (1) the utility functions are pre‑specified, limiting their ability to capture complex, heterogeneous traveler behavior, and (2) the Independence of Irrelevant Alternatives (IIA) assumption, which leads to unrealistic substitution patterns when routes share overlapping links. Recent hybrid approaches that combine discrete‑choice models with machine‑learning components improve predictive performance but typically rely on a pre‑generated candidate path set, making them difficult to scale to large networks, and they struggle to encode spatial network characteristics effectively.

The authors address these challenges by embedding a depth‑wise graph convolutional network (DGCN) within the RL formulation. The model retains the classic linear systematic utility component (e.g., travel time, distance, cost) for economic interpretability, while a non‑linear residual term generated by the GNN captures link‑level interactions and cross‑effects across the network. This residual term is added to the log‑probability of each link transition, effectively relaxing the IIA property without imposing strong distributional assumptions. An exogenous trade‑off parameter λ is introduced in the loss function to balance the contribution of the interpretable systematic utility against the data‑driven residual, allowing practitioners to control the degree of interpretability versus predictive accuracy.

Mathematically, the authors prove that a suitably designed GNN can represent multiple cross‑effect patterns, thereby providing a theoretical justification for the IIA relaxation. The model remains link‑based, preserving the key advantage of RL that it does not require explicit enumeration of the full choice set, which is critical for scalability in city‑wide networks.

Empirical validation is performed on a one‑day, large‑scale vehicle trajectory dataset collected in Tokyo. Raw GPS traces are matched to the road network to generate link‑level choice sequences. The authors compare ResDGCN‑RL against several benchmarks: standard MNL, C‑Logit, Path‑Size Logit, the recent ResLogit hybrid, and pure machine‑learning baselines such as Random Forest and deep neural networks. Evaluation metrics include Top‑1 prediction accuracy, log‑likelihood, and visual analysis of substitution patterns between overlapping routes. Results show that ResDGCN‑RL consistently outperforms all baselines, achieving higher accuracy and better calibrated substitution effects. Sensitivity analysis on λ demonstrates that modest values preserve most of the systematic utility’s interpretability while still delivering substantial gains in predictive performance.

The paper also discusses practical considerations. Training the GNN component incurs higher computational cost and requires careful hyper‑parameter tuning, but the authors argue that the benefits in accuracy and behavioral realism justify this overhead for large‑scale applications. Limitations include the need for substantial labeled trajectory data and the challenge of extending the approach to multimodal networks or real‑time settings.

Future research directions proposed are: (1) extending the framework to multimodal transportation systems, (2) incorporating real‑time traffic information into the GNN to enable dynamic route‑choice prediction, and (3) integrating Bayesian inference to quantify uncertainty in the learned residual terms.

In summary, this work makes three key contributions: (i) it proposes the first GNN‑based hybrid link‑based route‑choice model, (ii) it demonstrates theoretically and empirically that the model relaxes the IIA assumption and captures complex path substitution patterns, and (iii) it introduces a controllable trade‑off between economic interpretability and predictive power. The findings suggest that combining classic discrete‑choice theory with modern graph‑learning techniques offers a promising path forward for high‑fidelity, scalable transportation modeling.