OBE: Outlier by Example

Outlier detection in large datasets is an important problem. There are several recent approaches that employ very reasonable definitions of an outlier. However, a fundamental issue is that the notion of which objects are outliers typically varies between users or, even, datasets. In this paper, we present a novel solution to this problem, by bringing users into the loop. Our OBE (Outlier By Example) system is, to the best of our knowledge, the first that allows users to give some examples of what they consider as outliers. Then, it can directly incorporate a small number of such examples to successfully discover the hidden concept and spot further objects that exhibit the same “outlier-ness” as the examples. We describe the key design decisions and algorithms in building such a system and demonstrate on both real and synthetic datasets that OBE can indeed discover outliers that match the users’ intentions.