Discriminative features for identifying and interpreting outliers

摘要: We consider the problem of outlier detection and interpretation. While most existing studies focus on first problem, we simultaneously address equally important challenge propose an algorithm that uncovers outliers in subspaces reduced dimensionality which they are well discriminated from regular objects while at same time retaining natural local structure original data to ensure quality explanation. Our takes a mathematically appealing approach spectral graph embedding theory show it achieves globally optimal solution for objective subspace learning. By using number real-world datasets, demonstrate its performance not only w.r.t. rate but also discriminative human-interpretable features. This is exploit features both interpretation, leading better understanding how why hidden exceptional.