Indecision-Aware Deep Active Anomaly Detection

Abstract

This work introduces the $\mathrm {AE\!-\!AAD}$ algorithm, for Active Anomaly Detection through Auto-Encoders. Differently from pure unsupervised approaches, the algorithm has the possibility to improve output quality by incorporating the knowledge encoded by query answers. Specifically, we train an autoencoder-based architecture to directly reconstruct normally labelled and unlabelled examples and to maximize the difference between anomalously labelled examples and their reconstruction. Being the method aware of the target associated with both normal and abnormal data, we can introduce a notion of indecision which quantifies the maximum amount of deviation of a specific instance from its possible target reconstructions. Thus, our method is able to better discern between instances that are badly reconstructed because they comply with the anomalous target reconstruction from those that do not really conform to normal behavior. We perform experiments to clarify the behavior of the method and to compare performances with those of alternative anomaly detectors. Experimental results show that our method is able to exploit queries to improve the quality of the anomaly detection and also to ameliorate performances over other active anomaly detection proposals.