Model Construction

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.

ST-GAT Single Edge:

• Considers all edge types as the same.
• Utilizes a Graph Attention (GAT) layer to aggregate neighboring node features in a linear manner which enhances traditional Graph Convolutional Networks (GCN) with self-attention mechanisms.
• Employs two Long Short-Term Memory (LSTM) layers for capturing temporal characteristics.
• Employs a linear output layer for generating traffic speed forecasts.

ST-GAT Edge Type:

• Segregates graph signals based on edge types to discern their importance.
• Applies a GAT layer independently to each edge type for unique feature embeddings.
• Utilizes two LSTM layers for capturing temporal dependencies.
• Employs a linear output layer for generating traffic speed forecasts.

Graph Descriptions:

Graph	Node Features	Edge Types Included	Edge Types Learned
Graph1_SingleEdge	Speed	1	Not Learned
Graph2_SingleEdge	Speed	1,2	Not Learned
Graph3_SingleEdge	Speed	1,2,3	Not Learned
Graph4_SingleEdge	Speed, Lanes, Day of Week, Hour of Day	1	Not Learned
Graph5_SingleEdge	Speed, Lanes, Day of Week, Hour of Day	1,2	Not Learned
Graph6_SingleEdge	Speed, Lanes, Day of Week, Hour of Day	1,2,3	Not Learned
Graph1_EdgeType	Speed	1	Learned
Graph2_EdgeType	Speed	1,2	Learned
Graph3_EdgeType	Speed	1,2,3	Learned
Graph4_EdgeType	Speed, Lanes, Day of Week, Hour of Day	1	Learned
Graph5_EdgeType	Speed, Lanes, Day of Week, Hour of Day	1,2	Learned
Graph6_EdgeType	Speed, Lanes, Day of Week, Hour of Day	1,2,3	Learned