Diffusion Convolution Graph Attention Network for Spatial-Temporal Prediction

Abstract

Spatial-temporal prediction is widely used in railway, climate, smart city or other fields. In complex spatial-temporal data prediction, it is necessary to establish the corresponding temporal and spatial correlation model. Currently the main problems of spatial-temporal prediction focus on two aspects: firstly, the relevance and complexity of spatial data; secondly, the inherent difficulty of long-term prediction. In order to cope with these challenges, this paper propose a diffusion convolution graph attention network model to effectively capture the dependence of temporal and spatial. Specifically, we first use the bidirectional random walks to extract the correlation of local spatial dependence on the graph, then use the attention mechanism to capture the global spatial dependence. Finally, to deal with the difficulty of long-term prediction, the convolution Long Short-Term Memory (LSTM) network and the autoregressive component are used to capture the long-term pattern of the predicted data. The model is evaluated on three real large-scale spatial-temporal datasets. Results have proved it is effective compared with the advanced baseline model.