Extended B-ALIF: Improving Anomaly Detection with Human Feedback

Anomaly Detection is a task in engineering aiming at identifying deviations from expected patterns in data. Data-driven approaches have emerged in past recent years due to the fact that a model of complex system may be hard or impossible to be derived in many scenarios. Moreover, unsupervised approaches have been particularly appealing for practitioners and scientists given the typical unavailability of tagged data. Such approaches are often integrated in frameworks, like Decision Support Systems, that assist domain experts and operators in the monitoring task. Human presence, by providing a limited amount of feedback, can be leveraged as a valuable source of information to iteratively enhance detection performance. In this work we introduce Extended B-ALIF, a framework designed to incrementally select and integrate expert feedback into the Extended Isolation Forest anomaly detection model. This study extends Bayesian Active Learning Isolation Forest (B-ALIF), which originally proposed the same theoretical principles for another anomaly detection model, the Isolation Forest.