The article introduces a novel approach to Kriging, a method of estimating attributes of unseen locations based on neighboring data. The proposed method, called Contrastive- Prototypical self-supervised learning for Kriging (KCP), considers not only neighboring but also non-neighboring data. The model uses a neighboring contrastive module and a prototypical module to distinguish between useful and misleading data. An adaptive augmentation module is used to encourage data diversity. The model demonstrated superior performance in experiments, with a 6% improvement compared to previous methods.
Publication date: 23 Jan 2024
Project Page: https://github.com/bonaldli/KCP
Paper: https://arxiv.org/pdf/2401.12681
