Multivariate Traffic Demand Prediction via 2D Spectral Learning and Global Spatial Optimization

Changlu Chen, Yanbin Liu, Ling Chen, Chengqi Zhang

Abstract: Traffic demand forecasting is of critical importance to the development of Intelligent Transportation System (ITS). Existing works employ various deep learning models to capture the complex spatial-temporal correlations in the time-evolving traffic data. However, it has been proven that deep neural networks tend to learn low-frequency variations of inputs, leaving the high-frequency information underexplored. To overcome this spectral bias issue, we design a novel embedded 2D spectral learning framework. Firstly, a well-devised spectral embedding function is employed to encapsulate both the low-frequency and high-frequency signals of the multivariate input. This function also induces a shift-invariant kernel to maintain good distance metrics. Secondly, we model the temporal variations and multivariate feature interactions as two effective dimensions in the frequency domain. The 2D Fourier transform is directly applied along both dimensions followed by Fourier domain representation learning to extract more intrinsic patterns for traffic forecasting. Moreover, existing works employ either local aggregation or pairwise matching mechanism when modeling spatial correlations, which failed to consider the global structure of the traffic network. We propose to formulate spatial feature learning as an optimal transport problem to globally optimize the citywide spatial correlations and promote the overall performance of traffic. Extensive experiments are conducted on two real-world benchmarks with state-of-the-art results, which corroborate the effectiveness of embedded 2D spectral learning and global spatial optimization.