Hierarchically Structured Transformer Networks for Fine-Grained Spatial Event Forecasting

Spatial event forecasting is challenging and crucial for urban sensing scenarios, which is beneficial for a wide spectrum of spatial-temporal mining applications, ranging from traffic management, public safety, to environment policy making. In spite of significant progress has been made to solve spatial-temporal prediction problem, most existing deep learning based methods based on a coarse-grained spatial setting and the success of such methods largely relies on data sufficiency. In many real-world applications, predicting events with a fine-grained spatial resolution do play a critical role to provide high discernibility of spatial-temporal data distributions. However, in such cases, applying existing methods will result in weak performance since they may not well capture the quality spatial-temporal representations when training triple instances are highly imbalanced across locations and time. To tackle this challenge, we develop a hierarchically structured Spatial-Temporal ransformer network (STtrans) which leverages a main embedding space to capture the inter-dependencies across time and space for alleviating the data imbalance issue. In our STtrans framework, the first-stage transformer module discriminates different types of region and time-wise relations. To make the latent spatial-temporal representations be reflective of the relational structure between categories, we further develop a cross-category fusion transformer network to endow STtrans with the capability to preserve the semantic signals in a fully dynamic manner. Finally, an adversarial training strategy is introduced to yield a robust spatial-temporal learning under data imbalance. Extensive experiments on real-world imbalanced spatial-temporal datasets from NYC and Chicago demonstrate the superiority of our method over various state-of-the-art baselines.