增強型異常檢測：使用超過100個模型的綜合集成方法

Abstract

本研究提出並描述了一種創新性的集成方法，旨在提高各種異常數據集中的檢測和預測能力。所提出的框架結合了多種元模型的優勢，利用監督和非監督學習技術來實現卓越的性能。該方法的有效性通過六個不同數據集進行了嚴格評估：troponin、annthyroid、satellite、mammography、speech 和 letter。 首先，研究重點放在 troponin 數據集上，其中集成方法在主要不良心臟事件（MACE）的預測準確性方面顯示出顯著改進，超過了傳統指標如 HEART 分數。為了驗證所提出方法的普遍性和穩健性，我們進一步在 Odds 資料庫中選擇了五個廣泛使用的數據集進行測試。這些數據集因其多樣的特徵而被選中，為異常檢測算法提供了全面的評估基礎。 所提出的方法使用了總共110個模型，其性能與六個異常檢測數據集中的十種最先進（SOTA）方法進行了比較。結果一致顯示，所提出的方法在所有數據集上都優於傳統集成技術和SOTA模型，並在 AUC 和 AUPRC 分數上取得更高的成績。例如，在檢測異常值方面，所提出的方法顯著超過了單個檢測器、完整集成和其他基於表示學習的算法。 比較包括使用一組全面的排名指標進行的徹底評估，如 Kendall's Tau、Spearman's Rank Correlation、Normalized Discounted Cumulative Gain (NDCG) 和 Mean Squared Error (MSE)。此外，研究還探討了元模型和基礎模型的適用性和最佳數量。詳細實驗突顯了模型多樣性的重要性，SHAP（SHapley Additive exPlanations）值提供了關於單個模型貢獻的洞見。集成中的每個模型捕捉數據的不同方面，從而提升整體預測性能。 研究結果強調了所提出的集成方法在分類和異常檢測任務中的有效性。該研究為開發更可靠和更準確的預測模型做出了貢獻，強調了全面評估模型性能指標的必要性。未來的研究應集中於優化計算效率、解決潛在的過擬合問題，以及在不同環境中驗證該方法以確保更廣泛的適用性。這項工作代表了預測建模的重大進展，對臨床決策和其他關鍵應用具有重要影響。

This study is proposed, described, and demonstrated an innovative ensemble method designed to enhance detection and prediction capabilities across various abnormal datasets.The proposed framework combines the strengths of multiple meta-models, leveraging both supervised and unsupervised learning techniques to achieve superior performance. The method's efficacy is rigorously evaluated using six diverse datasets: troponin, annthyroid, satellite, mammography, speech, and letter. Initially, the focus is on the troponin dataset, where the ensemble method demonstrates significant improvements in predictive accuracy for major adverse cardiac events (MACE), surpassing traditional metrics like the HEART score. To validate the generalizability and robustness of the proposed method, we further test it on five widely-used datasets from the Odds repository. These datasets were selected for their diverse characteristics, providing a comprehensive evaluation foundation for abnormal detection algorithms. The proposed method employs a total of 110 models, and its performance is compared against ten state-of-the-art (SOTA) methods across six abnormal detection datasets. Results consistently show that the proposed method outperforms traditional ensemble techniques and SOTA models, achieving higher AUC and AUPRC scores across all datasets. For instance, the proposed method significantly surpasses individual detectors, full ensembles, and other representation learning-based algorithms in detecting abnormalities. The comparison includes a thorough assessment using a comprehensive set of ranking metrics, such as Kendall's Tau, Spearman's Rank Correlation, Normalized Discounted Cumulative Gain (NDCG), and Mean Squared Error (MSE). Additionally, the study explores the suitability and optimal number of meta-models and base models. Detailed experiments highlight the importance of model diversity, with SHAP (SHapley Additive exPlanations) values providing insights into individual model contributions. Each model within the ensemble captures distinct aspects of the data, enhancing overall predictive performance. The findings underscore the effectiveness of the proposed ensemble method in both classification and abnormal detection tasks. The study contributes to the development of more reliable and accurate predictive models, emphasizing the need for a comprehensive evaluation of model performance metrics. Future research should focus on optimizing computational efficiency, addressing potential overfitting, and validating the method across diverse environments to ensure broader applicability. This work represents a significant advancement in predictive modeling, with implications for clinical decision-making and other critical applications.