Abstract: To tackle the issue of missing traffic data in urban road network, this paper proposes a road network traffic data imputation method based on the spatial-temporal matrix factorization (namely STMF). Based on the spatiotemporal properties of the road network, this study first processes the multidimensional traffic data into a 2D matrix, then decomposes it into a spatial feature matrix and a temporal feature matrix and reconstructs the incomplete traffic data matrix by a low-rank approximation to achieve the basic repair of the missing data. Then graph Laplacian (GL) and gated recurrent network (GRN) are used as spatial and temporal regularizers respectively, to further mine spatial topology information and temporal dependencies of the urban road network and enhance the accuracy of matrix reconstruction and missing value imputation. Finally, the Los Angeles Traffic Speed dataset (Metr-LA) and Guangzhou Traffic dataset (Guangzhou-D) are used to compare the performance of STMF model with benchmark models such as GAIN, BGCP, BTMF, LRTC-TNN and HaLRTC. The experimental results show that, compared with the benchmark model, the proposed STMF model based on temporal matrix decomposition can better adapt to different missing scenarios and different missing rates, and the performance of missing data repair is more robust.

Key words: intelligent transportation, data imputation, matrix factorization, traffic data, graph Laplacian, gated recurrent network