Graph Attention LSTM: A Spatiotemporal Approach for Traffic Flow Forecasting
Tianqi Zhang, Ge Guo
Abstract
This article investigates a traffic-flow forecasting problem based on long–short term memory (LSTM), an artificial recurrent neural network architecture used in deep learning. By representing the road network as an unweighted directed graph, the traffic flow prediction problem becomes how to capture the spatiotemporal dependencies among nodes in the graph. We present a novel graph-attention LSTM structure, which leverages the strength of the graph-attention mechanism for non-Euclidean structured data modeling and that of the LSTM cell for time-series modeling. To demonstrate the effectiveness of the proposed method, two real-world data sets are evaluated. The results show that our model performs better than existing baselines.