Adaptive graph generation based on generalized pagerank graph neural network for traffic flow forecasting

Xiaoyu Guo, Xiangyuan Kong, Weiwei Xing, Xiang Wei, Jian Zhang, Wei Lu

Abstract: Traffic flow forecasting is a typical multivariate time series problem that has applications in intelligent transportation systems. It requires the modeling of complicated spatial-temporal dependencies and essential uncertainty regarding a road network and traffic conditions. Recently, some studies have improved their models without prespecified graphs by constructing adaptive matrices or learnable node embedding dictionaries; however, they omitted the semantic correlations among distant regions. In this paper, we propose an adaptive generalized PageRank graph neural network (AGP-GNN) for traffic flow forecasting, which jointly models spatial, temporal, and semantic correlations to adaptively generate hidden graph structures. Specifically, the AGP-GNN mainly includes two key components: 1) an adaptive generalized PageRank (AGP) layer, which dynamically assigns different edge weights to reflect the different correlations between the pairwise nodes; and 2) a relative position-based temporal attention (RPTA) layer, which models the complex correlations among different time steps. Moreover, we design a distance and temporal encoding (DTE) approach to incorporate geographic and temporal information. Experimental results obtained on two real-world datasets demonstrate the effectiveness of the AGP-GNN.