Enhancing few-sample spatio-temporal prediction via relational fusion-based hypergraph neural network

Abstract

Spatio-temporal prediction is a pivotal service for smart city applications, such as traffic and air quality prediction. Deep learning models are widely employed for this task, but the effectiveness of existing methods heavily depends on large amounts of data from urban sensors. However, in the early stages of smart city development, data scarcity poses a significant challenge due to the limited data collected from newly deployed sensors. Moreover, transferring data from other resource-rich cities is typically infeasible because of strict privacy policies. To address these challenges, we propose a relational fusion-based hypergraph neural network (RFHGN) for few-sample spatio-temporal prediction. RFHGN is trained directly on limited data within a city, exploiting multiple spatial correlations and hierarchical temporal dependencies to enrich spatio-temporal representations. Specifically, to enhance spatial expressiveness, we design a high-order spatial relation-aware learning module with an adaptive time-varying hypergraph structure. This structure is learned by integrating observational data and is iteratively updated during training, enabling the capture of dynamic high-order interactions. By combining these interactions with pairwise spatial representations, we derive mixed-order spatial representations. To reduce potential redundancy, we introduce a regularized independence loss to ensure the independence of pairwise and high-order spatial representations. Additionally, to effectively capture temporal dependencies at micro and macro levels, we develop a hierarchical temporal relation-aware learning module. Extensive experiments on three spatio-temporal prediction tasks: traffic flow, traffic speed, and air quality prediction demonstrate that RFHGN outperforms state-of-the-art baselines.