The rapid development of urbanization has facilitated the process of urban planning, e.g., new commercial and residential areas, which results in a surge of the local traffic demand. In this paper, we study the 'plan-based urban traffic estimation' problem. Effective urban plan estimation is of great significance to mitigate the operational vulnerability of a transportation system, such as traffic congestion and accidents. Solving this problem is challenging due to the complex spatio-temporal dependencies preserved in the various urban data. In this paper, we propose a novel urban traffic estimation framework, UTest, which can provide traffic estimations in consecutive time slots via spatio-temporal graph learning. UTest aims to provide accurate estimation results automatically, thus human planners can finally adjust machine-generated plans before deploying them. The proposed UTest adopts and advances the graph learning structure through a few novel ideas: (1) modeling various travel demands as the spatio-temporal graphs to build the spatio-temporal dependencies explicitly, (2) integrating graph convolution neural networks to learn the local spatial dependencies along the underlying road networks, (3) employing multi-head self-attention mechanism to learn the temporal dependencies of the traffic across different time slots. We conduct extensive experiments on two real-world spatio-temporal datasets. Experimental results demonstrate that UTest outperforms major baselines in estimation accuracy under various settings and can produce more meaningful estimation results.
