Non-Neighbors Also Matter to Kriging: A New Contrastive-Prototypical Learning

TL;DR

The paper targets Kriging under sparse sensor deployment by introducing a self-supervised framework, KCP, that leverages both neighboring and non-neighboring information. It combines a neighboring contrast module to align target embeddings with nearby nodes and a prototypical head to refine and recycle information from non-neighbors, guided by an adaptive augmentation strategy. Across three real-world datasets, KCP achieves consistent improvements of approximately 3–6% over strong baselines, demonstrating robust transferability and resilience to noise. This work shows that contrastive-prototypical SSL can enhance spatiotemporal interpolation by learning general, robust representations suitable for inductive Kriging in dynamic graphs.

Abstract

Kriging aims at estimating the attributes of unsampled geo-locations from observations in the spatial vicinity or physical connections, which helps mitigate skewed monitoring caused by under-deployed sensors. Existing works assume that neighbors' information offers the basis for estimating the attributes of the unobserved target while ignoring non-neighbors. However, non-neighbors could also offer constructive information, and neighbors could also be misleading. To this end, we propose ``Contrastive-Prototypical'' self-supervised learning for Kriging (KCP) to refine valuable information from neighbors and recycle the one from non-neighbors. As a pre-trained paradigm, we conduct the Kriging task from a new perspective of representation: we aim to first learn robust and general representations and then recover attributes from representations. A neighboring contrastive module is designed that coarsely learns the representations by narrowing the representation distance between the target and its neighbors while pushing away the non-neighbors. In parallel, a prototypical module is introduced to identify similar representations via exchanged prediction, thus refining the misleading neighbors and recycling the useful non-neighbors from the neighboring contrast component. As a result, not all the neighbors and some of the non-neighbors will be used to infer the target. To encourage the two modules above to learn general and robust representations, we design an adaptive augmentation module that incorporates data-driven attribute augmentation and centrality-based topology augmentation over the spatiotemporal Kriging graph data. Extensive experiments on real-world datasets demonstrate the superior performance of KCP compared to its peers with 6% improvements and exceptional transferability and robustness. The code is available at https://github.com/bonaldli/KCP

Figures (9)

• Figure 1: When looking through the node representations in the embedding space, spatial neighbors do not always appear nearby and non-neighbors may also be close to the target.
• Figure 2: The framework of the proposed KCP
• Figure 3: Embeddings of learned representations on PeMS. According to the ground truth, similar time series are clustered into the same class (same color).
• Figure 4: (a) Robustness and (b) Transferability.
• Figure 5: (a) Neighbors can be misleading: Baseline (IGNNK) uses misleading patterns from neighbors #2 and #3, who are close but less similar; (b) Non-neighbors can be constructive: KCP dynamically utilizes the patterns from the most similar nodes at each time $t$, e.g., at $t_0$, #23 is used since similarity = 0.95: some are neighbors (green), and some are non-neighbors (blue). Non-neighbors #45, #62, #170 used twice.