Applying Anomaly Pattern Score for Outlier Detection

Abstract

Outlier detection is an important sub-field of data mining and studied intensively by researchers in the past decades. For neighborhood-based outlier detection methods like KNN and LOF, different settings in the number of neighbors (indicated by a parameter k) would greatly affect the model's performance. Thereby, there are some recent studies which focus on identifying the optimal value of k by analyzing the global or local structure of the dataset. But, we argue that neighborhood-based outlier detection model could obtain an improvement in performance without parameter tuning. In this paper, from a novel angle of view, we adopt a uniform sampling strategy to generate a series of local proximity graphs and propose a new adaptive outlier detection model named anomaly pattern score which does not rely on the k tuning. In addition, the theoretical analysis of the effectiveness of the proposed model is conducted as well. The extensive experiments on both synthetic and real-world datasets show that the proposed model outperforms the state-of-the-art algorithms on most datasets.