Learning traffic patterns from road network data

Paper Title:

Heterogeneous Graph Neural Networks for Data-driven Traffic Assignment

Published on:

19 October 2023

Primary Category:

Machine Learning

Paper Authors:

Tong Liu,

Hadi Meidani

Bullets

Key Details

•

A heterogeneous graph neural network is proposed to model road networks and traffic flow

•

Virtual links between origins and destinations enable effective feature propagation

•

The model incorporates flow conservation laws to improve training

•

It generalizes well to new network topologies and link capacities

•

The approach enables data-driven traffic analysis without needing sensor data

Explore the topics in this paper

graph neural networks

machine learning

origin-destination flows

traffic prediction

transportation networks

AI generated summary

Learning traffic patterns from road network data

This paper proposes a new graph neural network model to analyze traffic patterns and estimate traffic flow in road networks, using only the network topology and origin-destination demand data. The model can capture complex spatial relationships and is generalizable to new network structures. Through message passing and attention mechanisms, it learns inherent traffic flow properties like flow conservation at nodes. Experiments demonstrate high accuracy and rapid training convergence compared to other models.