VOS: A new outlier detection model using virtual graph.

Outlier detection has been well studied due to its wide applications in both academia and industry, among which graph-based methods have drawn extensive attention in recent years because of their robust expressiveness for various types of datasets. Combining the local information with the implicit connections in the graph representation of the original dataset, in this study, we propose a new outlier detection model named Virtual Outlier Score (VOS). The proposed model constructs a similarity graph by using the top-k similar neighbors of each object, and introduces a virtual node coupling with a collection of virtual edges to generate a k-virtual graph. A tailored Markov random walk process is then performed on the strongly connected virtual graph under the principle that a potential outlier should get more weight to be visited. After reaching equilibrium, the stationary distribution vector is utilized to deduce the virtual outlier score. Furthermore, we provide a theoretical analysis of the proposed VOS model. Experiments on both synthetic and real-world datasets showed that the proposed model obtains an improvement over eight state-of-the-art algorithms under the measures of ROC AUC as well as the outlier discovery curve.