SubRank: Ranking Local Outliers in Projections of High-Dimensional Spaces
Outlier mining has become an increasingly urgent issue in the KDD process, since it may be the case that finding exceptional events is more interesting than searching for common patterns. These outliers are most relevant to be found for instance in fraud detection processes. Unfortunately, existing approaches do not take into account that increasing dimensionality leads to a novel understanding of locality. Objects have to be investigated locally in different projections of the original space to overcome a crucial problem: The outlier property might be occluded by the shear number of dimensions. Being aware of this breach, in the course of my diploma thesis, I developed a novel, effective method, SubRank, to rank objects which are outliers only in some subspaces. Finally, it gives a concise explanation of the composition of the ranking itself.
|Authors:||Schiffer M., Müller E., Seidl T.|
|Published in:||Datenbank-Spektrum Vol. 9 Issue 29|
|Publisher:||Springer - Heidelberg, Germany|
|Forschungsgebiet:||Data Analysis and Knowledge Extraction|