Scholarlink

Performance anomalies in database systems degrade service quality, making efficient diagnosis essential. However, existing methods rely on large amounts of labeled data, which is costly to obtain in practice. Active learning tackles this by selectively labeling samples. Despite this, most active learning query strategies prioritize samples with high uncertainty while neglecting those with high discrepancy, hindering the rapid and accurate establishment of decision boundaries. To address these challenges, we propose ActiveDiag, an active learning-based framework for anomaly diagnosis in database systems. It incorporates a Discrepancy and Uncertainty Dynamic Fusion (DUDF) query strategy and a Transformer-based Siamese Network (TSN) classifier to enhance diagnostic performance. Specifically, DUDF considers both discrepancy and uncertainty, dynamically fusing them at different query stages. Additionally, TSN combines the powerful feature representation capability of transformers and the precise similarity measurement capability of siamese networks. Furthermore, we introduce a threshold-constrained cold start strategy to mitigate inaccurate uncertainty estimation of samples during the early query stage, particularly when certain predefined categories are absent from the labeled dataset. Extensive experiments show ActiveDiag achieves superior diagnostic performance, achieving a Macro-F1 of 0.8820 and a Micro-F1 of 0.9070 with 50 expert-labeled samples, outperforming the best previous framework by 9.31% and 9.69%.

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

ActiveDiag: Dynamic Fusion of Discrepancy and Uncertainty in Active Learning for Database Anomaly Diagnosis

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Peize Yuan,
  • Zixuan Li,
  • Xiyue Gao,
  • Mingzhe Wang,
  • Hui Li,
  • Yanguo Peng,
  • Yaofeng Tu,
  • Jiangtao Cui

摘要

Performance anomalies in database systems degrade service quality, making efficient diagnosis essential. However, existing methods rely on large amounts of labeled data, which is costly to obtain in practice. Active learning tackles this by selectively labeling samples. Despite this, most active learning query strategies prioritize samples with high uncertainty while neglecting those with high discrepancy, hindering the rapid and accurate establishment of decision boundaries. To address these challenges, we propose ActiveDiag, an active learning-based framework for anomaly diagnosis in database systems. It incorporates a Discrepancy and Uncertainty Dynamic Fusion (DUDF) query strategy and a Transformer-based Siamese Network (TSN) classifier to enhance diagnostic performance. Specifically, DUDF considers both discrepancy and uncertainty, dynamically fusing them at different query stages. Additionally, TSN combines the powerful feature representation capability of transformers and the precise similarity measurement capability of siamese networks. Furthermore, we introduce a threshold-constrained cold start strategy to mitigate inaccurate uncertainty estimation of samples during the early query stage, particularly when certain predefined categories are absent from the labeled dataset. Extensive experiments show ActiveDiag achieves superior diagnostic performance, achieving a Macro-F1 of 0.8820 and a Micro-F1 of 0.9070 with 50 expert-labeled samples, outperforming the best previous framework by 9.31% and 9.69%.