Robust Anomaly Detection from Partially Observed Anomalies with Augmented Classes

Abstract

Anomaly detection is becoming increasingly ubiquitous in the society of data mining. Prominent anomaly detection works have achieved great success in theory and practice. However, they cannot handle the generalized semi-supervised scenario where there are only a handful of labeled anomalies, and plentiful unlabeled data that may bring in some instances of augmented anomaly classes but which are hard to be sampled. To solve this new problem, we propose a method called ACAD (Augmented Classes Anomaly Detection), which consists of three components. ACAD firstly suggests an augmented anomaly class discovery module that connects the isolation score and the similarity score to excavate the instances of hidden anomaly classes from unlabeled data accurately. ACAD then uses a specific cluster approach to compute useful similarity scores to separate reliable anomalous and normal instances among unlabeled data, respectively. ACAD finally builds a robust anomaly detector based on mined examples, successfully performing anomaly detection from partially observed anomalies with augmented classes. A series of empirical studies show that our algorithm remarkably outperforms state of the art on almost twenty datasets.