Enhancing Graph Neural Networks in Large-scale Traffic Incident Analysis with Concurrency Hypothesis

TL;DR

The Concurrency Prior (CP) method is proposed, a powerful approach designed to enhance the predictive capabilities of general Graph Neural Network models in semi-supervised traffic incident prediction tasks, which allows GNNs to incorporate concurrent incident information, as mentioned in the hypothesis, via tokenization with negligible extra parameters.

Abstract

Despite recent progress in reducing road fatalities, the persistently high rate of traffic-related deaths highlights the necessity for improved safety interventions. Leveraging large-scale graph-based nationwide road network data across 49 states in the USA, our study first posits the Concurrency Hypothesis from intuitive observations, suggesting a significant likelihood of incidents occurring at neighboring nodes within the road network. To quantify this phenomenon, we introduce two novel metrics, Average Neighbor Crash Density (ANCD) and Average Neighbor Crash Continuity (ANCC), and subsequently employ them in statistical tests to validate the hypothesis rigorously. Building upon this foundation, we propose the Concurrency Prior (CP) method, a powerful approach designed to enhance the predictive capabilities of general Graph Neural Network (GNN) models in semi-supervised traffic incident prediction tasks. Our method allows GNNs to incorporate concurrent incident information, as mentioned in the hypothesis, via tokenization with negligible extra parameters. The extensive experiments, utilizing real-world data across states and cities in the USA, demonstrate that integrating CP into 12 state-of-the-art GNN architectures leads to significant improvements, with gains ranging from 3% to 13% in F1 score and 1.3% to 9% in AUC metrics. The code is publicly available at https://github.com/xiwenc1/Incident-GNN-CP.

Figures (8)

• Figure 1: The state-wise Average Neighbor Crash Density (ANCD) for Top: negative nodes (i.e. nodes without incident records) and Bottom: positive nodes (i.e. nodes with incident records) when $k=1$. For a specific class of nodes (i.e. positive/negative nodes), a deeper color denotes a higher density of their neighbor nodes have incident. It is observed that, within the same state, the neighbor nodes of positive nodes often exhibit a higher crash density than those of negative, which supports our Concurrency Hypothesis.
• Figure 2: The graph-based data is obtained from real-world road networks.
• Figure 3: The statistics of the graph-based datasets on 49 states.
• Figure 4: Left: The problem is formulated in a single large graph. Nodes' labels are known if the node is marked in colors, i.e., red (positive)/ blue (negative). The nodes with question marks are expected to be predicted. Middle: In the training phase, we keep all unknown nodes with the uncertain token 'o', and in each iteration, we also randomly mask some nodes with known labels to 'o' to mimic the prediction process. Right: In the inference phase, we only keep to nodes to be predicted with the uncertain token 'o'. Top: Imposing concurrency prior to the neural network.
• Figure 5: The statistics of ANCD and ANCC computed for the available 49 states.

Theorems & Definitions (1)

• Theorem 1