Outlier Detection Using Diverse Neighborhood Graphs

Abstract

Owing to its wide applications in both industry and academia, a large number of new approaches are emerging every year in the field of outlier detection. Among which, neighborhood-based approaches are adopted by a great number of researchers and they still represent the mainstream in the field. However, how to determine appropriate local information from the definition of neighbors is an arduous problem which still has no widely accepted solution. In this study, we propose a new outlier detection model utilizing multiple neighborhood graphs, each of which is based on changed neighbors to capture various local information from different perspectives. An outlier score for each object is then deduced by performing random walk on the predefined graphs. Experiments on ten real-world datasets suggested that the proposed model could obtain promising results compared with four state-of-the-art algorithms by the measure of ROC AUC and precision at n.