|Urban Traffic Forecasting has recently seen a lot of research activity as it entails a compelling combination of multivariate temporal data with geo-spatial dependencies between multiple data collection sensors. Current top approaches to this task tend to use costly spatio-temporal pipelines, where the model complexities typically have linear dependency on the time-series length and quadratic on the number of nodes. In this paper, we propose a number of steps to dramatically improve the runtime efficiency of the traffic forecasting solutions. First, we use a temporal pooling stack prior to spatial processing to effectively eliminate the time dimension before applying the spatial components. This removes the linear dependency of the model on the length of the time series. Second, we construct learnable graph pooling blocks inside the spatial stack which progressively reduce the size of the graph and facilitate better data flow between far away nodes. Experimental results on the standard METR-LA and PEMSBAY benchmarks show that the proposed approach yields significant inference and training speedups of up to x5 in the 1-hour prediction task and x27 in the 24-hour prediction task, while keeping or surpassing the state-of-the-art results. Our findings call into question the need for time-consuming spatio-temporal processing blocks, used in many of latest solutions for the traffic forecasting task.
*** Title, author list and abstract as seen in the Camera-Ready version of the paper that was provided to Conference Committee. Small changes that may have occurred during processing by Springer may not appear in this window.