Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Multi-Class Traffic Assignment using Multi-View Heterogeneous Graph Attention Networks

Tong Liu, Hadi Meidani

TL;DR

The paper tackles efficient multi-class traffic assignment on large networks by introducing a multi-view heterogeneous graph attention network (M-HetGAT) that assigns a dedicated view to each vehicle class, includes origin-destination links, and enforces flow conservation as a loss regularizer. The model uses intra- and inter-class attention to capture both within-class and cross-class interactions, and predicts edge-level metrics such as the flow-capacity ratio to obtain link flows and utilization. Through extensive experiments on Sioux Falls, EMA, and Anaheim networks under user-equilibrium and system-optimal scenarios, M-HetGAT outperforms baseline multi-view GNNs in MAE/RMSE, exhibits high correlation with ground truth, and generalizes to altered topologies with robust conservation behavior. The approach provides a scalable surrogate capable of accelerating TAP computations and can be extended to dynamic demand, weather, and other external factors, offering practical benefits for real-time transportation management.

Abstract

Solving traffic assignment problem for large networks is computationally challenging when conventional optimization-based methods are used. In our research, we develop an innovative surrogate model for a traffic assignment when multi-class vehicles are involved. We do so by employing heterogeneous graph neural networks which use a multiple-view graph attention mechanism tailored to different vehicle classes, along with additional links connecting origin-destination pairs. We also integrate the node-based flow conservation law into the loss function. As a result, our model adheres to flow conservation while delivering highly accurate predictions for link flows and utilization ratios. Through numerical experiments conducted on urban transportation networks, we demonstrate that our model surpasses traditional neural network approaches in convergence speed and predictive accuracy in both user equilibrium and system optimal versions of traffic assignment.

Multi-Class Traffic Assignment using Multi-View Heterogeneous Graph Attention Networks

TL;DR

The paper tackles efficient multi-class traffic assignment on large networks by introducing a multi-view heterogeneous graph attention network (M-HetGAT) that assigns a dedicated view to each vehicle class, includes origin-destination links, and enforces flow conservation as a loss regularizer. The model uses intra- and inter-class attention to capture both within-class and cross-class interactions, and predicts edge-level metrics such as the flow-capacity ratio to obtain link flows and utilization. Through extensive experiments on Sioux Falls, EMA, and Anaheim networks under user-equilibrium and system-optimal scenarios, M-HetGAT outperforms baseline multi-view GNNs in MAE/RMSE, exhibits high correlation with ground truth, and generalizes to altered topologies with robust conservation behavior. The approach provides a scalable surrogate capable of accelerating TAP computations and can be extended to dynamic demand, weather, and other external factors, offering practical benefits for real-time transportation management.

Abstract

Solving traffic assignment problem for large networks is computationally challenging when conventional optimization-based methods are used. In our research, we develop an innovative surrogate model for a traffic assignment when multi-class vehicles are involved. We do so by employing heterogeneous graph neural networks which use a multiple-view graph attention mechanism tailored to different vehicle classes, along with additional links connecting origin-destination pairs. We also integrate the node-based flow conservation law into the loss function. As a result, our model adheres to flow conservation while delivering highly accurate predictions for link flows and utilization ratios. Through numerical experiments conducted on urban transportation networks, we demonstrate that our model surpasses traditional neural network approaches in convergence speed and predictive accuracy in both user equilibrium and system optimal versions of traffic assignment.
Paper Structure (20 sections, 20 equations, 5 figures, 4 tables)

This paper contains 20 sections, 20 equations, 5 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Literature
  3. Technical Background
  4. Multi-Class Vehicle Traffic Assignment Problem
  5. Graph Neural Network
  6. Multi-View Heterogeneous Graph Attention Network for Traffic Assignment
  7. Graph Construction & Feature Preprocessing
  8. Multi-View Heterogeneous Graph Attention Network
  9. Graph Edge Prediction
  10. Numerical Experiments
  11. Experiments on Urban Transportation Networks
  12. Characteristics of networks
  13. Experiment setup
  14. Numerical results
  15. Experiments on Urban Transportation Networks with Altered Topologies
  16. ...and 5 more sections

Figures (5)

  • Figure 1: A schematic of the multi-view heterogeneous graph attention network for traffic assignment. The proposed model consists of three parts: (1) graph construction and feature preprocessing module, (2) multi-view graph attention, and (3) multi-view graph fusion. The flow-capacity ratio and link flow of each link are calculated using the source node feature, destination node feature, and normalized edge feature.
  • Figure 2: Comparison of predicted versus ground truth link flows in the Sioux Falls network, using four different models.
  • Figure 3: The illustration of urban transportation networks with the removed links shown in red.
  • Figure 4: Comparison of node-level flow conservation residue in the East Massachusetts network. Four models are included: GAT, GCN, GraphSAGE, and ours
  • Figure 5: Ablation study of different variants of M-HetGAT. Three variations of M-HetGAT are included: "w/o link feat": model maintaining OD link but removing road link features; "w/o OD link": model removing OD link; "w/o Intra view": model removing intra-view message passing; "w/o Multi-head": model removing multi-head attention mechanism; "w/o conservation": model removing flow conservation in the loss function