Finding outliers at multiple scales

Outlier detection targets those exceptional data whose pattern is rare and lie in low density regions. In this paper, under the assumption of complete spatial randomness inside clusters, we propose an MDV (Multi-scale Deviation of the Volume) approach to identifying outliers. In addition to assigning an outlier score for each object, it directly outputs a crisp outlier set. It also offers a plot showing the data structure in every object's vicinity, which is useful in explaining why it may be outlying. Finally, the effectiveness of MDV is demonstrated with both artificial and real datasets.