Spatiotemporal urban data (STUD) displays complex correlational patterns. Extensive advanced techniques have been designed to capture these patterns for effective forecasting. However, because STUD is often massive in scale, practitioners need to strike a balance between effectiveness and efficiency by choosing computationally efficient models. An alternative paradigm called MLP-Mixer has the potential for both simplicity and effectiveness. Taking inspiration from its success in other domains, we propose an adapted version, named NexuSQN, for STUD forecast at scale. We identify the challenges faced when directly applying MLP-Mixers as series- and window-wise multivaluedness and propose the ST-contextualization to distinguish between spatial and temporal patterns. Experimental results surprisingly demonstrate that MLP-Mixers with ST-contextualization can rival SOTA performance when tested on several urban benchmarks. Furthermore, it was deployed in a collaborative urban congestion project with Baidu, specifically evaluating its ability to forecast traffic states in megacities like Beijing and Shanghai. Our findings contribute to the exploration of simple yet effective models for real-world STUD forecasting.
View on arXiv