Identifying Environmentally Hazardous Road Segments Using Graph Neural Networks

Road segments with extreme traffic congestion can cause harm to the environment through increased CO2 emissions, respiratory health issues as an effect of pollution, decreased productivity due to time spent in traffic, and increased car accidents. Congestion on certain roads surrounded by areas with normal traffic flow can be even more dangerous due to their disruption of regular traffic dynamics and greater vehicle emissions. Machine Learning (ML) models such as Graph Neural Networks (GNNs) have become popular for identifying unusual patterns. Out of the three types of classification, node-level methods have been extensively explored in unsupervised learning, while edge-level classification methods have received less attention. Since our task involves edge-level classification, to bridge the gap we convert the edge-level problem to node-level using line graphs and identify road segments with abnormal congestion in the Santa Barbara road network using GNNs. Through this, we are able to identify areas for improvement of traffic flow in order to improve the environment and public health/safety. Expanding models to include edge-level classification widens ML's scope to assist in effective analysis and decision-making. GNNs outperform most ML models for identifying these congested road segments in an unsupervised setting because they can process multiple features in data, process non-Euclidean relationships, analyze relationships between data points, and more. Our model using GNNs is successfully able to identify congested roads in historical traffic data and identifies roads that are unnaturally congested with respect to their adjacent roads that models using only statistical thresholds of traffic data cannot.