LSTM variants meet graph neural networks for road speed prediction

Traffic flow prediction is a fundamental issue in smart cities and plays an important role in urban traffic planning and management. An accurate predictive model can help individuals make reliable travel plans and choose optimal routes while efficiently helping administrators maintain traffic order. Road speed prediction, which is a sub-task of traffic flow forecasting, is challenging due to the complicated spatial dependencies characterizing road networks and dynamic temporal traffic patterns. Given the power of recurrent neural networks (RNNs) in learning temporal relations and graph neural networks (GNNs) in integrating graph-structured and node-attributed features, in this paper, we design a novel graph LSTM (GLSTM) framework to capture spatial-temporal representations in road speed forecasting. More specifically, we first present a temporal directed attributed graph to model complex traffic flow. Then, to take advantage of the structure properties and graph features, we employ a message-passing mechanism for feature aggregation and updating. Finally, we further implement several variants of LSTMs with a GN block under the encoder-decoder framework to model spatial-temporal dependencies. The experiments show that our proposed model is able to fully utilize both the road latent graph structure and traffic speed to forecast the road state during future periods. The results on two real-world datasets show that our GLSTM can outperform state-of-the-art baseline methods by up to 32.8% in terms of MAE, 43.2% in terms of MAPE and 23.1% in terms of RMSE.