PlugSI: Plug-and-Play Test-Time Graph Adaptation for Spatial Interpolation

TL;DR

PlugSI addresses the challenge of adapting spatial interpolation (SI) models to unseen and evolving test-time graphs in sensor networks. It introduces two plug-and-play components: an Unknown Topology Adapter (UTA) that refines the test-time graph for each small batch and a Temporal Balance Adapter (TBA) that preserves historical consistency while adapting to new data, guided by a self-supervised objective enforcing spatial-temporal consistency. The method comprises a Virtual Uncertainty Scorer (VUS) and an Adjacency Refiner (AR) within UTA, and a memory-based update rule within TBA, achieving refined adjacency matrices ĤA_t used by fixed pre-trained SI models. Across four real-world datasets and seven backbones, PlugSI yields significant improvements (e.g., notable reductions in MAE) with robust performance under streaming data and varying virtual-node patterns, while maintaining lightweight computation and scalability.

Abstract

With the rapid advancement of IoT and edge computing, sensor networks have become indispensable, driving the need for large-scale sensor deployment. However, the high deployment cost hinders their scalability. To tackle the issues, Spatial Interpolation (SI) introduces virtual sensors to infer readings from observed sensors, leveraging graph structure. However, current graph-based SI methods rely on pre-trained models, lack adaptation to larger and unseen graphs at test-time, and overlook test data utilization. To address these issues, we propose PlugSI, a plug-and-play framework that refines test-time graph through two key innovations. First, we design an Unknown Topology Adapter (UTA) that adapts to the new graph structure of each small-batch at test-time, enhancing the generalization of SI pre-trained models. Second, we introduce a Temporal Balance Adapter (TBA) that maintains a stable historical consensus to guide UTA adaptation and prevent drifting caused by noise in the current batch. Empirically, extensive experiments demonstrate PlugSI can be seamlessly integrated into existing graph-based SI methods and provide significant improvement (e.g., a 10.81% reduction in MAE).

Figures (5)

• Figure 1: Overall insight of our study. Fig. 1(a) depicts the unknown graph topology at test-time, and Fig. 2(c) shows it should be refined before inference. Meanwhile, Fig. 1(b) highlights current batch may overwrite historical refinement information. So Fig. 1(d) thinks the information of historical and current information should be balanced.
• Figure 2: The overall framework of PlugSI. At a high level, it refines the local topology of virtual sensors in each small-batch $\mathbf{B}_t$, yielding a refined adjacency matrix $\mathbf{\hat{A}}_t$. PlugSI comprises two components: UTA, where VUS estimates topological uncertainty $\mathbf{S}_t$ for virtual sensors and AR derives the adaptation $\Delta\mathbf{A}_t$ to form $\mathbf{\hat{A}}_t$; and TBA, which integrates the current adaptation with accumulated historical information to generate updated guidance $\mathbf{A}'_t$ for the current batch and ensure stable online adaptation.
• Figure 3: Ablation study for IGNNK and KITS on METR-LA dataset. As observed, each component is effective.
• Figure 4: Impact of number of information balance coefficient $\alpha$, regularization weight $\lambda$ and small-batch size $B$ for IGNNK and KITS on METR-LA dataset.
• Figure 5: Experimental results under different proportions of virtual nodes for IGNNK and KITS on the METR-LA dataset.