Although active learning is an established solution to reduce labeling costs for classification problems, its effectiveness in an outlier detection context has yet to be studied. In this paper, we offer encouraging experimental results that suggest that the benefits of active learning, with a particular instance selection strategy, may indeed apply to outlier detection. In addition, we offer an analysis of how the performance gains brought by active learning vary with the sizes of the initial training set and of the set of instances added by active learning.

This is joint work with Nathalie Japkowicz and Colin Bellinger