In a world where automobiles are ubiquitous, traffic congestion remains a persistent challenge. Existing solutions often fall short in predicting traffic flow effectively, leaving commuters trapped in congestion. Our innovative approach leverages graph neural networks (GNNs) to revolutionize traffic prediction by integrating time series and geospatial analysis methodologies. By focusing on recent traffic sensor data in San Diego, our endeavor aims to forecast traffic flow accurately, empowering drivers with timely information to navigate congested routes more efficiently.

Take a look into our data collection process.

Explore the diverse representations of highway networks through distinct graph structures, each tailored with unique edge connections and node features. Dive into the dynamics of traffic flow across various time frames within a unified structural framework.

For traffic speed prediction, we adopt the Spatial-Temporal Graph Attention Network (ST-GAT) proposed by Zhang, Yu, and Liu (2019). ST-GAT efficiently integrates spatial relationships and temporal variations within the network, enabling precise traffic forecasting. Two variants of ST-GAT are introduced in our graph implementations: ST-GAT Single Edge and ST-GAT Edge Type.

Check out our model performance.