Convolution-aware networks for random missing traffic data imputation

Abstract

The integrity of traffic data is fundamental to alleviating the challenges in urban cities by computing. However, traffic data often exhibits a random missing characteristic due to sensor failure or network packet loss. Existing methods endowed too much prior knowledge on random missing data, such as data decay over time or data distribution correlation analysis. Thus, there is an urgent need for a data-driven and efficient traffic data interpolation method to assist downstream urban computing. Therefore, this paper proposes a fully convolutional spatial-temporal graph neural network (FC-STGNN) for traffic data imputation. Specifically, we apply a temporal convolutional network (TCN) to extract temporal features. Due to the dilated causal convolutions, it is possible to extract temporal features across time nodes, effectively alleviating the impact of data loss at a certain moment. Furthermore, we design a graph convolutional network (GCN) with residual connections to aggregate traffic data between adjacent road segments in the road network. Combining these two components enables spatiotemporal modeling of traffic data in data-missing environments. Finally, we conduct experiments on two real-world traffic datasets. The experiments demonstrate that our proposed method outperforms most baseline methods and owns a modest computational cost.