Spatial-Temporal Traffic Data Imputation via Graph Attention Convolutional Network

High-quality traffic data is crucial for intelligent transportation system and its data-driven applications. However, data missing is common in collecting real-world traffic datasets due to various factors. Thus, imputing missing values by extracting traffic characteristics becomes an essential task. By using conventional convolutional neural network layers or focusing on standalone road sections, existing imputation methods cannot model the non-Euclidean spatial correlations of complex traffic networks. To address this challenge, we propose a graph attention convolutional network (GACN), a novel model for traffic data imputation. Specifically, the model follows an encoder-decoder structure and incorporates graph attention mechanism to learn spatial correlation of the traffic data collected by adjacent sensors on traffic graph. Temporal convolutional layers are stacked to extract relations in time-series after graph attention layers. Through comprehensive case studies on the dataset from the Caltrans performance measurement system (PeMS), we demonstrate that the proposed GACN consistently outperforms other baselines and has steady performance in extreme missing rate scenarios.