因此，本研究的目標是通過訓練和比較五種較具有代表性的非監督式異常偵測方法所建立之模型效能，為未來研究人員在利用事件紀錄檔進行相關非監督式異常檢測時建立可靠的參考基礎。為了確保結果的可靠性，本研究使用了三個分別記錄了不同的報銷流程的真實世界事件紀錄檔以進行模型訓練並回答研究問題。此外，本研究基於事件紀錄檔常見的三個基本元素（時間、資源、活動），定義了七個異常情境，以便比較不同模型之間的性能表現。通過實驗結果的評估和比較後，我們發現局部異常因子偵測方法（ local outlier factor, LOF）是以報銷流程的事件紀錄檔進行異常偵測時，最適用之非監督式異常檢測方法。;Information systems generate valuable log data in modern business processes crucial in various investigations. Anomaly detection using this data is essential for identifying outliers, detecting fraud, and supporting process improvement and risk management. Among the different categories of anomaly detection, unsupervised anomaly detection techniques stand out for their practicality in real-world applications, thanks to their minimal requirements. Previous research on anomaly detection in business processes has predominantly concentrated on utilizing specific anomaly detection techniques, which lack comparison between models and are often conducted without employing public datasets. Therefore, this research aims to establish a reliable foundation for future researchers interested in utilizing log data for unsupervised anomaly detection in business processes.
This is achieved by training and comparing five representative unsupervised anomaly detection algorithms. To ensure the reliability and robustness of the results, three real live event log datasets, capturing distinct reimbursement processes, are utilized to address the research questions. Additionally, seven anomaly scenarios, based on the three essential elements (time, resource, activity) commonly found in event logs, are defined to facilitate the comparison of performance between different models. Through evaluation and comparison, it is revealed that the local outlier factor (LOF) is the most suitable unsupervised algorithm for detecting anomalies in reimbursement process event logs.
