Learning Dynamic Graph Embedding for Traffic Flow Forecasting: A Graph Self-Attentive Method

Abstract

The Traffic flow forecasting (TFF) problem is essential to modern intelligent transport systems (ITS). Massive flow data from contemporary transport systems have put forward the challenge of effectively capturing both the latent spatial patterns and the temporal dynamics of traffic flow data, when an ITS is doing forecasting. To cope with this challenge, we introduce a novel spatial-temporal graph self-attentive model (STGSA) for short-term traffic flow forecasting. Our model learns graph-level spatial embedding using graph self-attention layers with Gumbel-Softmax technique, and the temporal embedding leveraging RNN cells integrated with Gated Recurrent Units. We evaluate the effectiveness of our proposed method on the traffic flow data of Langfang, China throughout the year of 2014, and it outperforms most of the state-of-the-art baselines.