Long-term Traffic Prediction Using Time-varying Adjacency Mask and Self-Smoothing Regularization

Abstract

Long-term traffic prediction is essential for pre-control of traffic departments, which allows traffic dispatchers to make earlier decisions than short-term traffic prediction. This task is extremely challenging mainly due to the difficulty of obtaining accurate spatial dependency at different time periods and weak correlation between predicted values and historical data for largest time step. Existing methods either use the same adjacency matrix at every moment or recompute a different adjacency matrix at every moment, which may introduce incorrect neighbors to the target node. Moreover, many previous methods either obtain the predicted values step by step, which causes error propagation problem, or obtain the predicted values for each step independently, which loses the correlation information between the multi-step predicted values. In this paper, a Time-varying Adjacency Mask is proposed to correct the spatial dependence which makes spatial dependence different but highly similar at each moment. Besides, a Self-Smoothing Regularization is proposed to establish the relationship between the predicted values of adjacent time slices and to restrict their differential values. Extensive experiments including traditional long-term traffic speed prediction, time-phased speed prediction and rush hour speed prediction are conducted on two real-world datasets, experimental results show the superior performance of our proposed model.