Scholarlink

<p>Forecasting complex human health outcomes like college student fitness demands models that handle nonlinear relationships while grasping the evolving links among behavior indicators over time. Existing machine learning methods often stumble under assumptions of linearity, static setups, or feature independence. These fail to capture real-world dynamics mirrored in wearable sensor data. To break free, we introduce a fresh modeling approach: Dynamic Graph-Kernelized Regression (DGKR). This end-to-end unified framework delivers a full "dimensional upgrade" to tackle three core hurdles. It evolves the Health Behavior Network (HBN) from a static snapshot into a time-varying graph sequence. It lifts regression via the kernel trick into an infinite-dimensional Reproducing Kernel Hilbert Space (RKHS) for arbitrary complexity. And it weaves in robustness through personalized self-attention and quantile regression for adaptive handling and uncertainty measures. We applied DGKR to a real college health dataset (FAW) and several general time-series regression tasks from the UEA &amp; UCR archive. Broad experiments show DGKR tops prediction accuracy across the board, statistically outpacing 11 strong baselines spanning diverse categories, including XGBoost, LSTM, and state-of-the-art dynamic graph networks.</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

Unveiling the dynamics of student physique: a time-varying graph kernel approach for collegiate fitness assessment from wearable data

  • 期刊论文

  • 发表日期：

    2026年3月4日

查看全文
  • Dilixiati Diliya,
  • Nan Wen

摘要

Forecasting complex human health outcomes like college student fitness demands models that handle nonlinear relationships while grasping the evolving links among behavior indicators over time. Existing machine learning methods often stumble under assumptions of linearity, static setups, or feature independence. These fail to capture real-world dynamics mirrored in wearable sensor data. To break free, we introduce a fresh modeling approach: Dynamic Graph-Kernelized Regression (DGKR). This end-to-end unified framework delivers a full "dimensional upgrade" to tackle three core hurdles. It evolves the Health Behavior Network (HBN) from a static snapshot into a time-varying graph sequence. It lifts regression via the kernel trick into an infinite-dimensional Reproducing Kernel Hilbert Space (RKHS) for arbitrary complexity. And it weaves in robustness through personalized self-attention and quantile regression for adaptive handling and uncertainty measures. We applied DGKR to a real college health dataset (FAW) and several general time-series regression tasks from the UEA & UCR archive. Broad experiments show DGKR tops prediction accuracy across the board, statistically outpacing 11 strong baselines spanning diverse categories, including XGBoost, LSTM, and state-of-the-art dynamic graph networks.