Scholarlink

<p>Unsupervised anomaly detection of sequential patterns is attractive in various industries. Compared to methods that rely on prior model assumptions, model-free methods are more adaptive to different applications, while are more challenging. Some popular methods quantify the absolute error between the actual and expected frequencies of a pattern. However, they are insensitive to new patterns. In this paper, we propose a three-way anomaly detection of sequential pattern (3WADSP) method to address this issue. First, we define the error between a&#xa0;pattern’s actual frequency relative to the expected one as anomaly metric. The detection performance is effectively improved due to the high sensitivity of the new&#xa0;metric. Second, we construct two trisecting-acting-outcome models for estimating expected frequency, and capturing the symptoms before the occurrence and disappearance of anomalous patterns, respectively. The former provides better interpretability, and the later brings stronger diversity. Finally, we design a smoothing technique by averaging the anomaly degrees of overlapped patterns. The ambiguity caused by the overlapping is eliminated. 3WADSP not only extends the methodology of three-way decisions, but also provides new tools and perspectives for applications such as medical and fault diagnosis.</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

Three-way unsupervised anomaly detection of sequential patterns

  • 期刊论文

  • 发表日期：

    2025年1月20日

查看全文
  • Gong-Suo Chen,
  • Tirapot Chandarasupsang,
  • Zhi-Heng Zhang,
  • Xiang-Bing Zhou,
  • Wu Deng,
  • Annop Tananchana,
  • Lei Mu,
  • Fan Min

摘要

Unsupervised anomaly detection of sequential patterns is attractive in various industries. Compared to methods that rely on prior model assumptions, model-free methods are more adaptive to different applications, while are more challenging. Some popular methods quantify the absolute error between the actual and expected frequencies of a pattern. However, they are insensitive to new patterns. In this paper, we propose a three-way anomaly detection of sequential pattern (3WADSP) method to address this issue. First, we define the error between a pattern’s actual frequency relative to the expected one as anomaly metric. The detection performance is effectively improved due to the high sensitivity of the new metric. Second, we construct two trisecting-acting-outcome models for estimating expected frequency, and capturing the symptoms before the occurrence and disappearance of anomalous patterns, respectively. The former provides better interpretability, and the later brings stronger diversity. Finally, we design a smoothing technique by averaging the anomaly degrees of overlapped patterns. The ambiguity caused by the overlapping is eliminated. 3WADSP not only extends the methodology of three-way decisions, but also provides new tools and perspectives for applications such as medical and fault diagnosis.