Spatiotemporal Traffic Flow Modelling

Foundational work developing graph neural network approaches to citywide traffic prediction, with particular focus on robustness to sensor failure and data sparsity.

Overview

This completed project established the graph neural network framework that underpins several subsequent projects. The key innovation was a topology-aware attention mechanism that remains stable when up to 30% of sensor nodes report missing data — a common real-world condition that prior models handled poorly.

Key Findings

Spatiotemporal GNNs outperform LSTM baselines by 18% RMSE on sparse sensor networks
Transfer learning from dense to sparse sensor environments requires domain-adaptive pre-training, not simple fine-tuning
Prediction accuracy degrades gracefully up to 30% node failure; beyond this threshold performance drops sharply

Dataset

The project produced the AUS-Traffic-2020 dataset: 3 years of 5-minute interval flow counts from 847 sensors across Southeast Queensland, now publicly available under CC-BY 4.0.

Publications

J2, J3, C2, C3 in the publications list stem directly from this project.