A Robust Method for Detecting DB-Outliers from High Dimensional Datasets

Outlier detection is a popular technique that can be utilized in many modern applications like financial analysis and fraud detection. As data description becomes complex, operated datasets’ dimensionalities keep monotone increasing. However, current researches find that it is extremely difficult to pick out outliers directly from high dimensional datasets owing to the curse of dimensionality. Moreover, general methods need some decisive parameters to be decided in advance. Such parameters usually have close connection with data distribution. Therefore, users always have no idea about proper parameters without identifying datasets beforehand. To address these problems, we introduce a method to discover exceptional objects that match users’ intentions in high dimensional datasets. Compared with determining proper parameters, users are more easily to provide some outlier examples including their intensions. We make good use of outlier examples, examine behaviors of projections of these examples and find an optimal subspace. The concept of DB(Distance-Based) Outliers is employed to detect outliers in the optimal subspace. Our proposed method is robust to tolerate noises or inconsistencies interfusing in outlier examples. Experiments operate on both synthetic and real datasets. The results show that our proposed approach is effective and efficient at detecting outliers corresponding with users ’intensions in high dimensional datasets.