MS2OD: outlier detection using minimum spanning tree and medoid selection
Jia Li, Jiangwei Li, Chenxu Wang, Fons J. Verbeek, Tanja Schultz, Hui Liu
Abstract
Abstract As an essential task in data mining, outlier detection identifies abnormal patterns in numerous applications, among which clustering-based outlier detection is one of the most popular methods for its effectiveness in detecting cluster-related outliers, especially in medical applications. This article presents an advanced method to extract cluster-based outliers by employing a scaled minimum spanning tree (MST) data structure and a new medoid selection method: 1. we compute a scaled MST and iteratively cut the current longest edge to obtain clusters; 2. we apply a new medoid selection method, considering the noise effect to improve the quality of cluster-based outlier identification. The experimental results on real-world data, including extensive medical corpora and other semantically meaningful datasets, demonstrate the wide applicability and outperforming metrics of the proposed method.