All Times EDT
Keywords: Convex Hull Peeling, One-Class Methods, Support Vector Data Description
Outlier detection is important for preprocessing data and/or for detecting anomalous observations. Numerous outlier detection methods have been proposed in the fields of statistics, machine learning, and data mining. Some outlier detection methods are based on distance measures and require estimation of the sample covariance matrix. We propose here a flexible framework to detect multiple outliers that does not require covariance estimation, is well-suited to high-dimensional datasets with a high percentage of outliers, and is robust to parameter specification. We evaluate our framework using both synthetic and benchmark data sets, showing that it works well in high dimensions and performs better than benchmark methods, especially when there is a higher percentage of outliers. The proposed methodology can be adapted to different type of data including unstructured data, such as anomalous text categorization.