Di-GraphGAN: An enhanced adversarial learning framework for accurate spatial-temporal traffic forecasting under data missing scenarios

Abstract

Nowadays, various disturbances in urban transportation data acquisition/processing/storage lead to the inevitable data missing problem, which undermines the valuable traffic information and greatly threats the reliability of existing benchmark traffic prediction models. Inspired from the powerful generative learning ability of GANs, we propose an integrated spatiotemporal Data imputation Graph Attention Generative Adversarial Networks (Di-GraphGAN) for accurate and efficient spatial-temporal traffic forecasting under data missing scenarios. Specifically, we first propose a traffic data imputation module named DI-LSTM, which adopts the architecture of LSTM Network with an extra Time Damping unit to accurately estimating the missing values. Then, we facilitate Di-GraphGAN with an original developed Task-Efficient Graph Attention Networks (TE-GAT) for better graph representation learning and a Temporal Contextual Attention (TCA) mechanism to capture the dynamic spatiotemporal traffic patterns. Finally, extensive evaluations are conducted on two real-world traffic speed datasets from China, demonstrating that Di-GraphGAN achieves state-of-the-art performance in both traffic forecasting and spatiotemporal data imputation tasks.