Exploring memory-facilitated graph neural networks for traffic forecasting

Traffic forecasting is essential for optimising intelligent transport systems (ITS), directly impacting traffic congestion, accident reduction, and transportation efficiency. With the success of deep learning, methods like graph neural networks and Transformers have been applied to traffic prediction. However, these approaches face significant challenges due to the unpredictable nature of traffic. Main issues include the distribution shift in traffic data, lack of uncertainty estimation in predictions, and the risk that traffic sensors might record data compromised by noise.

This thesis presents innovations in graph neural networks, enhanced with memory components to address these challenges. We introduce a novel memory-facilitated framework for distribution shift adaptation in Chapter~\ref{ST-align}, improving accuracy through a dual-component strategy that adapts to shifting distributions in real-world datasets. For uncertainty estimation, the memory-augmented conditional neural process (MemCNP) model in Chapter~\ref{memcnp} leverages neural processes to capture intricate spatio-temporal patterns and generate corresponding uncertainty estimations, thereby improving performance across various traffic contexts. Additionally, a robust noisy label learning framework in Chapter~\ref{nll} enhances model resilience against sensor noise, validated on benchmark datasets.

Overall, integrating memory mechanisms into graph neural networks significantly enhances the robustness, accuracy, and generalisability of traffic forecasting models. These methodologies aim to address fundamental challenges in traffic data analysis, contributing new insights to the application of ITS. Rigorous testing and real-world applications confirm the models' potential to transform traffic management and planning, demonstrating their value in both academic research and practical implementation.