Abstract
Traffic data plays an essential role in Intelligent Transportation
Systems (ITS) and offers numerous advantages, including efficient
traffic control and system performance improvement. However, due to the
scarcity of data collection systems, missing data in traffic datasets is
inevitable. Therefore, traffic data imputation becomes an essential
task. Graph Neural Network (GNN) is a type of neural network that
operates on graph-structured data and has shown potential in handling
traffic network related tasks such as traffic prediction and traffic
data imputation. In this paper, we contribute to the body of knowledge
with two aspects. First, we focus on traffic data imputation using
solely spatial information. Most of the studies in the literature
address spatio-temporal traffic data imputation, which is a distinct
task from our research. Second, Since most GNN models operates on node
features, we propose an approach to construct node features for nodes in
traffic networks, by leveraging available link flows. To investigate the
effectiveness of the proposed method, we implement two missing
scenarios, random missing (RM) and block missing (BM). We evaluate the
performance of the proposed method on three different sized real-world
networks: Sioux Falls, Anaheim, and Chicago. The evaluation results
demonstrate that GNN models outperform other baselines for most of the
missing patterns.