Scholarlink

<p>Structure learning aims to uncover the underlying dependencies or relationships between variables in a network. Traditional methods rely on statistical inference to reveal causal relationships or conditional dependencies. However, there is a growing need for innovative approaches that enhance network merging and improve both the accuracy and scalability of structure learning. In this study, we propose a novel method that improves network merging by combining multiple network structures based on their conditional dependencies, while leveraging domain knowledge to eliminate unrealistic dependencies, something that traditional methods often overlook. This innovative approach enhances both the accuracy and scalability of structure learning by prioritizing the most relevant dependencies and ensuring the structural integrity of the network. We evaluate this approach using environmental quality data collected from a commercial building in Winnipeg, Canada, to predict the likelihood of abnormal conditions. The results show our novel method outperforms the classical method by an average of 4.4% in accuracy. Furthermore, we improved the proposed method by integrating environmental data from neighboring locations, which enhanced the accuracy of anomaly detection by an average of 5.4%. In conclusion, the comparative analysis demonstrates that the Bayesian network model, incorporating our proposed structure learning method, performs better in anomaly detection compared to classical structure learning approaches.</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

Enhanced anomaly detection through a Bayesian framework with a novel network merging structure learning approach

  • 期刊论文

  • 发表日期：

    2025年3月5日

查看全文
  • Ashani Wickramasinghe,
  • Saman Muthukumarana

摘要

Structure learning aims to uncover the underlying dependencies or relationships between variables in a network. Traditional methods rely on statistical inference to reveal causal relationships or conditional dependencies. However, there is a growing need for innovative approaches that enhance network merging and improve both the accuracy and scalability of structure learning. In this study, we propose a novel method that improves network merging by combining multiple network structures based on their conditional dependencies, while leveraging domain knowledge to eliminate unrealistic dependencies, something that traditional methods often overlook. This innovative approach enhances both the accuracy and scalability of structure learning by prioritizing the most relevant dependencies and ensuring the structural integrity of the network. We evaluate this approach using environmental quality data collected from a commercial building in Winnipeg, Canada, to predict the likelihood of abnormal conditions. The results show our novel method outperforms the classical method by an average of 4.4% in accuracy. Furthermore, we improved the proposed method by integrating environmental data from neighboring locations, which enhanced the accuracy of anomaly detection by an average of 5.4%. In conclusion, the comparative analysis demonstrates that the Bayesian network model, incorporating our proposed structure learning method, performs better in anomaly detection compared to classical structure learning approaches.