Scholarlink

<p>Incomplete or missing sensor data pose a fundamental challenge in spatiotemporal computing, where reliable data recovery is essential for accurate modeling and inference. Existing graph-based imputation methods often overlook the varying informativeness of nodes across space and time, leading to biased or unstable reconstructions. In this paper, we propose an attribution-driven node calibration framework that adaptively adjusts node importance through a principled attribution analysis. By integrating gradient-based attribution into the spatiotemporal graph learning process, our framework quantifies each node’s contribution to the reconstruction objective and applies controlled perturbations to prevent dominance by noisy or redundant nodes. The resulting optimization balances informativeness and robustness, improving both numerical stability and interpretability. Extensive experiments on real-world traffic and environmental datasets show that our framework achieves superior imputation accuracy and robustness over state-of-the-art baselines. Beyond sensor networks, the proposed framework offers a general computational paradigm for learning under incomplete graph-structured data. Our code is available at: <a href="https://github.com/kaibo-Z/SAGO">https://github.com/kaibo-Z/SAGO</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Attribution-driven node calibration for spatiotemporal data recovery

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Kaibo Zhang,
  • Tong Jia

摘要

Incomplete or missing sensor data pose a fundamental challenge in spatiotemporal computing, where reliable data recovery is essential for accurate modeling and inference. Existing graph-based imputation methods often overlook the varying informativeness of nodes across space and time, leading to biased or unstable reconstructions. In this paper, we propose an attribution-driven node calibration framework that adaptively adjusts node importance through a principled attribution analysis. By integrating gradient-based attribution into the spatiotemporal graph learning process, our framework quantifies each node’s contribution to the reconstruction objective and applies controlled perturbations to prevent dominance by noisy or redundant nodes. The resulting optimization balances informativeness and robustness, improving both numerical stability and interpretability. Extensive experiments on real-world traffic and environmental datasets show that our framework achieves superior imputation accuracy and robustness over state-of-the-art baselines. Beyond sensor networks, the proposed framework offers a general computational paradigm for learning under incomplete graph-structured data. Our code is available at: https://github.com/kaibo-Z/SAGO.