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

Zhishuai Li, Yunhao Nie, Ziyue Li, Lei Bai, Yisheng Lv, Rui Zhao

Introduction

Spatial interpolation, or Kriging, is a fundamental task in geostatistics and environmental monitoring, used to estimate values at unsampled locations based on observed data from nearby sensors. Traditionally, these methods rely on the "Law of Neighborhood," which posits that physically closer nodes are more likely to share similar characteristics. However, this assumption often fails in complex urban or environmental systems. For instance, a sensor on a quiet residential street may be physically close to a major highway but exhibit vastly different traffic patterns. Conversely, two commercial districts miles apart might show near-identical temporal behaviors.

To address these limitations, we introduce KCP (Kriging Contrastive-Prototypical learning), a self-supervised framework that moves beyond raw spatial proximity. KCP learns to "refine" the influence of misleading neighbors and "recycle" constructive information from distant non-neighbors by mapping nodes into a learned embedding space based on functional similarity.

The KCP Framework

KCP operates on a Graph Neural Network (GNN) backbone, specifically GraphSAGE, to enable inductive learning. The framework is built upon three core components designed to improve representation quality and robustness.

1. Adaptive Data Augmentation

Unlike standard random masking, KCP employs a learning-based augmentation strategy. Using a Gumbel-Softmax operator, the model adaptively decides whether to mask specific features or entire nodes during training. Furthermore, "topological augmentation" is performed by dropping edges connected to high-centrality nodes. This forces the model to learn from diverse subsets of neighbors, preventing over-reliance on a few dominant nodes and improving generalization.

2. Neighboring Contrastive Module (NC)

The NC module serves as a "coarse" filter that respects traditional spatial physics. It employs contrastive learning to push the embeddings of physical neighbors closer together while pulling non-neighbors apart. This ensures that the learned representations retain the basic spatial structure of the network before finer adjustments are made.

3. Prototypical Head (PH)

The PH module acts as the "fine-grained" refiner. It clusters nodes into latent "prototypes" based on their actual time-series patterns rather than their coordinates. By utilizing the Sinkhorn optimal transport algorithm, the model identifies similar functional representations across the entire map. This allows KCP to "recycle" information from a distant node if it shares a prototype with the target location, effectively identifying "functional neighbors" that physical distance would overlook.

Experimental Evaluation

The KCP framework was evaluated on three diverse spatiotemporal datasets: traffic flow (PeMS), solar power output (NREL), and wind speed (Wind).

Performance and Robustness

KCP consistently outperformed state-of-the-art statistical and deep learning baselines, including OKriging, IGNNK, and KCN. On average, KCP achieved a 6% improvement in Mean Absolute Error (MAE). Beyond raw accuracy, the model demonstrated superior robustness to sensor noise. In simulations of random sensor failures, KCP’s performance degraded significantly more slowly than its competitors, highlighting the benefit of its global "recycling" mechanism.

Zero-Shot Transferability

A key advantage of KCP’s inductive nature is its transferability. In tests where the model was trained on traffic data from the PeMS-Bay area and tested on the METR-LA (Los Angeles) dataset without retraining, it maintained high accuracy. This suggests that KCP learns fundamental spatiotemporal principles that transcend specific geographical layouts.

Conclusion

KCP represents a paradigm shift in spatial interpolation by demonstrating that "representation speaks louder than neighbors." By combining local spatial constraints with global functional prototypes, the framework effectively navigates the pitfalls of misleading neighbors and the untapped potential of distant non-neighbors. While the computational complexity of the Sinkhorn algorithm presents a scaling challenge for massive networks, KCP sets a new benchmark for accuracy and robustness in kriging tasks.

Summary updated

BibTeX
@inproceedings{li2024non,
  author = {Zhishuai Li and Yunhao Nie and Ziyue Li and Lei Bai and Yisheng Lv and Rui Zhao},
  title = {{Non-Neighbors Also Matter to Kriging: A New Contrastive-Prototypical Learning}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2024},
  url = {https://openreview.net/forum?id=PpIPBOuxxi},
}
Figure