Scholarlink

<p>Spatio-temporal forecasting is critical in the traffic domain, where accurate predictions are essential for effective urban traffic management, planning, and simulation. Despite the importance of complete historical observations, missing values due to sensor failures, data transmission errors, and other issues are common, posing significant challenges to the accuracy and reliability of forecasting models. Existing methods often fail to systematically account for incomplete historical data, especially non-random data missing for extended periods. Fortunately, this study introduces the MissNet, a pre-training enhanced framework for spatio-temporal data forecasting in the presence of missing historical data. MissNet consists of a two-stage process: a pre-training stage where a data masking and recovering task is used to pre-train a backbone, and a fine-tuning stage where the pre-trained backbone, combined with a specially designed header, predicts future data incorporating spatio-temporal metadata as auxiliary information. Experimental results on real-world datasets demonstrate the effectiveness of MissNet in achieving stable and accurate predictions under various missing data scenarios.</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

MissNet: Leveraging Pre-trained Network for Spatio-temporal Forecasting with Missing Observations

  • 期刊论文

  • 发表日期：

    2025年5月23日

查看全文
  • Shen Fang,
  • Hongyan Liu,
  • Wei Hua

摘要

Spatio-temporal forecasting is critical in the traffic domain, where accurate predictions are essential for effective urban traffic management, planning, and simulation. Despite the importance of complete historical observations, missing values due to sensor failures, data transmission errors, and other issues are common, posing significant challenges to the accuracy and reliability of forecasting models. Existing methods often fail to systematically account for incomplete historical data, especially non-random data missing for extended periods. Fortunately, this study introduces the MissNet, a pre-training enhanced framework for spatio-temporal data forecasting in the presence of missing historical data. MissNet consists of a two-stage process: a pre-training stage where a data masking and recovering task is used to pre-train a backbone, and a fine-tuning stage where the pre-trained backbone, combined with a specially designed header, predicts future data incorporating spatio-temporal metadata as auxiliary information. Experimental results on real-world datasets demonstrate the effectiveness of MissNet in achieving stable and accurate predictions under various missing data scenarios.