Graph Neural Network Surrogate for Strategic Transport Planning

Introduction

Transportation is a cornerstone of modern society, with urban environments becoming increasingly complex. This complexity poses significant challenges for accurately modeling and simulating transportation systems. Surrogate models have emerged as a promising approach to address these challenges, offering easier application and sometimes even surpassing the performance of traditional models. This paper builds upon previous work by exploring advanced Graph Neural Network (GNN) architectures as surrogate models for strategic transport planning. Specifically, it compares Graph Convolution Networks (GCN) with Graph Attention Networks (GAT) and introduces a novel GAT variant (GATv3) to address over-smoothing issues. Additionally, a hybrid model combining GCN and GAT architectures is proposed. The study aims to expand the horizons of GNN-based surrogate modeling in strategic transport planning.

Literature Review

GNNs for Transport Planning

Graph Neural Networks (GNNs) have gained traction in transport planning due to their ability to capture intricate spatial dependencies within urban networks. Various studies have demonstrated the effectiveness of GNNs in predicting demand for micromobility modes, taxi trips, and short-term traffic forecasting. For instance, Lin et al. (2018) used a GCN model to forecast station-level hourly demand for a bike-sharing system in New York City, showing promising results. Similarly, Song et al. (2023) developed a GNN model to predict demand for a shared E-scooter system in Louisville, USA.

Surrogate Modelling in Transportation

Surrogate models approximate expensive functions and replicate input-output relationships, offering benefits such as cheaper evaluation costs and sometimes better accuracy. In traffic forecasting, Vlahogianni (2015) introduced surrogate modeling, showing that it can generate more accurate predictions in significantly reduced times. Ge et al. (2020) used surrogate models for dynamic user equilibrium traffic assignment, reducing computational times by up to 90% while maintaining lower errors. The concept of GNN-based surrogate modeling in transport emerged recently, with Narayanan et al. (2024) laying the groundwork by investigating the feasibility of developing a GCN-based surrogate for strategic transport planning.

Summary and Research Gaps

Existing literature shows that GNNs have potential for transport modeling, and surrogate models are computationally effective. However, the limitations of GCNs and the need for advanced GNN architectures necessitate further investigation. This study aims to address these gaps by exploring advanced GNN architectures and novel modeling setups.

Methodological Framework

This study builds on the framework established by Narayanan et al. (2024), involving data generation, simulation, data transformation, and GNN model training and evaluation. The study introduces additional data generation processes and further GNN models, implemented in Python using libraries like Scikit-Learn and PyTorch Geometric.

Synthetic Data Generation

To overcome the limitations of finite subnetwork generation, a synthetic data generation procedure is implemented. This algorithm creates random nodes within a square region, generates zonal information, and creates routes between zones. The synthetic dataset, consisting of 10,000 samples, is used as supplementary training data.

Graph Attention Networks

Graph Attention Networks (GAT) address the limitation of GCNs by applying attention mechanisms to differentiate between neighbors. The study introduces GATv3, a novel extension that uses a residual trick to enable deeper networks, addressing the over-smoothing problem.

Combined GAT and GCN

To address the over-squashing problem, a hybrid model combining GAT and GCN architectures is proposed. The model first uses GAT layers to determine zonal demand and then applies GCN layers to realistically apply the demand onto the network.

Experimental Setup

The study focuses on three distinct setups: (i) a classification setup to compare different GNNs, (ii) a hierarchical regression setup to combine classification and regression tasks, and (iii) a fine-grained classification setup to convert classification results into precise values. The impact of additional synthetic data is also explored.

Comparison of GCNII, GATv2, GATv3, and GCN+GATv3

The first experiment compares different GNNs based on a classification task. The best performing model is GATv3, followed closely by the GCN+GAT model and then the GCNII model. All three GNNs perform similarly and significantly better than the baselines.

Hierarchical Regression

The hierarchical regression setup aims to combine the stability of classification with the precision of regression. However, the results are suboptimal, indicating that deeper analysis is needed to understand the challenges and potential solutions in this specific modeling context.

Fine-Grained Classification

The fine-grained classification setup explores the stability of classification and proposes a method to convert probabilities into real values. The results show that fine-grained classification performs well and is more stable with respect to model hyperparameters than regression. The use of additional synthetic data significantly improves model performance.

Experimental Results

Evaluation of Different GNN Architectures

The GATv3 model performs best in the classification task, followed by the GCN+GAT model and the GCNII model. The hierarchical regression setup yields suboptimal results, indicating the need for further research.

Evaluation of Fine-Grained Classification

Fine-grained classification with additional synthetic data shows significant improvement in performance. The GCNII model with extra data performs best, followed closely by the GCN+GAT model. The results highlight the importance of additional training data for better predictions.

Conclusions

This study advances GNN-based surrogate modeling for strategic transport planning by exploring advanced GNN architectures and novel modeling setups. The findings highlight the effectiveness of the new GATv3 model in classification tasks and the potential of fine-grained classification with additional training data. The study provides valuable insights for future research and application in urban planning scenarios.