Decoupling Long- and Short-Term Patterns in Spatiotemporal Inference

TL;DR

This work tackles real-time spatiotemporal inference under sparse sensor networks by decoupling short-term and long-term pattern learning. It introduces DualSTN, a dual-branch architecture combining a Joint Spatiotemporal Graph Attention Network for short-term joint dependencies with a Skip Graph Gated Recurrent Unit for long-term trends, aided by inductive graph sampling and pseudo-nodes for unseen locations. The approach achieves state-of-the-art results on four public datasets while maintaining fewer parameters and offering interpretable attention weights that reveal dynamic dependencies. The findings demonstrate that separating and explicitly modeling short- and long-term spatiotemporal relations yields superior accuracy and generalization in environmental sensing tasks with practical implications for smart city deployments.

Abstract

Sensors are the key to environmental monitoring, which impart benefits to smart cities in many aspects, such as providing real-time air quality information to assist human decision-making. However, it is impractical to deploy massive sensors due to the expensive costs, resulting in sparse data collection. Therefore, how to get fine-grained data measurement has long been a pressing issue. In this paper, we aim to infer values at non-sensor locations based on observations from available sensors (termed spatiotemporal inference), where capturing spatiotemporal relationships among the data plays a critical role. Our investigations reveal two significant insights that have not been explored by previous works. Firstly, data exhibits distinct patterns at both long- and short-term temporal scales, which should be analyzed separately. Secondly, short-term patterns contain more delicate relations including those across spatial and temporal dimensions simultaneously, while long-term patterns involve high-level temporal trends. Based on these observations, we propose to decouple the modeling of short-term and long-term patterns. Specifically, we introduce a joint spatiotemporal graph attention network to learn the relations across space and time for short-term patterns. Furthermore, we propose a graph recurrent network with a time skip strategy to alleviate the gradient vanishing problem and model the long-term dependencies. Experimental results on four public real-world datasets demonstrate that our method effectively captures both long- and short-term relations, achieving state-of-the-art performance against existing methods.

Figures (8)

• Figure 1: (a) Examples of three types of spatiotemporal dependencies. (b) Geographical distribution of stations and unknown locations. (c) PM2.5 readings and wind condition. The curves represent approximated trends and the arrows are time skips for long-term learning. The gray string denotes the time of 16:00, 09/11/2017 where signals do not tally with location relations.
• Figure 2: The framework of the proposed DualSTN model which contains two components: SG-GRU for long-term learning and JST-GAT for short-term learning. Adj means adjacency matrix.
• Figure 3: Illustrations of graph sampling. For each iteration, we sample a subgraph and divide nodes into observed sensors and unknown locations randomly.
• Figure 4: Illustrations and information flows of $k$-nearest inverse distance weighting and the joint spatiotemporal attention. Only a limited number of flows for one node are drawn.
• Figure 5: Ablation studies. The DualSTN consistently achieves the best RMSE and MAE results against other variants.