Memory-augmented dynamic graph convolution networks for traffic data imputation with diverse missing patterns

Abstract

Missing data is an inevitable and ubiquitous problem for traffic data collection in intelligent transportation systems. Recent research has employed graph neural networks (GNNs) for spatiotemporal data imputation and achieved promising performance. However, there still exist two limitations to be addressed: first, existing approaches are generally limited in directly leveraging global spatiotemporal information from different nodes at different time; second, most of these approaches do not consider the unique characteristics of transportation systems or traffic data, including dynamic spatial dependencies and correlated missing patterns. To fill these research gaps, we propose a novel deep learning framework called Memory-augmented Dynamic Graph Convolution Networks (MDGCN) to impute missing traffic data. The model uses a recurrent layer to capture temporal information and a graph convolution layer to capture spatial information. To address the first research gap, we introduce an external memory network to store and share the global spatiotemporal information across the traffic network. For the second research gap, a graph structure estimation technique is proposed to learn dynamic spatial dependencies directly from traffic data. In addition, four types of missing patterns with various missing ratios are considered in model evaluation. Extensive experiments based on two public traffic speed datasets are conducted. The results show that our proposed model outperforms existing state-of-the-art deep learning approaches in all kinds of missing scenarios, and both the proposed external memory network and graph structure estimation technique contribute to the model performance. The model performance is competitive in most cases even without complete training data.