Multi-domain Active Learning for Semi-supervised Anomaly Detection

Active learning aims to ease the burden of collecting large amounts of annotated data by intelligently acquiring labels during the learning process that will be most helpful to learner. Current active learning approaches focus on learning from a single dataset. However, a common setting in practice requires simultaneously learning models from multiple datasets, where each dataset requires a separate learned model. This paper tackles the less-explored multi-domain active learning setting. We approach this from the perspective of multi-armed bandits and propose the active learning bandits (Alba) method, which uses bandit methods to both explore and exploit the usefulness of querying a label from different datasets in subsequent query rounds. We evaluate our approach on a benchmark of 7 datasets collected from a retail environment, in the context of a real-world use case of detecting anomalous resource usage. Alba outperforms existing active learning strategies, providing evidence that the standard active learning approaches are less suitable for the multi-domain setting.