Scholarlink

<p>We present a scalable graph-based framework for assessing and clustering student risk profiles using campus-wide behavioral datasets. Developed for university security systems, the approach identifies risk-aligned student communities through hierarchical behavioral analysis. Activity features—such as late-night library access or facility usage frequency—are enriched with weakly supervised learning to capture diverse risk signals. A geometric feature selection strategy removes weak indicators, yielding refined behavioral representations. These are mapped into a latent space where students are modeled as probabilistic distributions over abstract risk factors, enabling precise threat differentiation. A weighted similarity graph supports institution-scale community detection via multi-resolution clustering, uncovering shared behavioral patterns such as irregular attendance or unusual social interactions. A prediction module integrates individual profiles with community trends to generate targeted alerts, strengthening early-warning capabilities. Tested on over 50,000 students, the framework achieved 89% accuracy, demonstrating scalability, robustness, and compliance with privacy standards.</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

Huge-graph-based risk communities mining and prediction: a smart campus security protection system toward effective students management

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Ling Tan

摘要

We present a scalable graph-based framework for assessing and clustering student risk profiles using campus-wide behavioral datasets. Developed for university security systems, the approach identifies risk-aligned student communities through hierarchical behavioral analysis. Activity features—such as late-night library access or facility usage frequency—are enriched with weakly supervised learning to capture diverse risk signals. A geometric feature selection strategy removes weak indicators, yielding refined behavioral representations. These are mapped into a latent space where students are modeled as probabilistic distributions over abstract risk factors, enabling precise threat differentiation. A weighted similarity graph supports institution-scale community detection via multi-resolution clustering, uncovering shared behavioral patterns such as irregular attendance or unusual social interactions. A prediction module integrates individual profiles with community trends to generate targeted alerts, strengthening early-warning capabilities. Tested on over 50,000 students, the framework achieved 89% accuracy, demonstrating scalability, robustness, and compliance with privacy standards.