| Abstract: | Anomaly detection, which aims to identify unusual or uncommon behaviors in data, has many real-world applications. While numerous machine learning algorithms have been developed for anomaly detection, we often still need extensive engineering efforts to develop effective machine learning pipelines. Moreover, anomaly detectors can have high false positive rates because they are often unsupervised and trained on imbalanced data, which significantly impedes the practical use of anomaly detection algorithms. To bridge the gaps, in this thesis, we aim to provide a unified anomaly detection module to support automated anomaly detection in various real-world applications and an active anomaly detection algorithm to reduce the high false positive rates. It is challenging to achieve these two objectives because of several unanswered research questions in I/O interface design, detection algorithm interface design, and the difficulty of selecting the right sample for query in active anomaly detection. Through answering these questions, we devise a consistent I/O interface for different types of anomalies, a unified objected-oriented design of anomaly detection algorithms, and a novel active anomaly detection algorithm based on meta-learning. Extensive experiments show that the implemented anomaly detection module is effective and that the proposed active anomaly detection algorithm outperforms the existing base-lines. Part of the research outcomes has been integrated into the TODS package, a time-series anomaly detection system. The electronic version of this dissertation is accessible from https://hdl.handle.net/1969.1/198663 |
|---|