RSOD: Efficient Technique for Outlier Detection using Reverse Nearest Neighbors Statistics

Unsupervised learning techniques are popular in detecting outliers in various domains. Many parametric and non-parametric outlier detection approaches have been proposed over the last decades. The existing neighborhood-based non-parametric unsupervised approaches like LOF, symmetric neighborhood, LDOF are proven to be effective when outliers are in a region of variable density. However, these techniques wrongly treat an outlier point as inlier in certain scenarios (outlier located between a dense cluster and close to a sparse cluster). In this work, we address this problem by exploiting the information of k-nearest neighbors and reverse nearest neighbors efficiently. We conducted experiments with synthetic, and four real-world datasets, and our proposed technique outperforms popular symmetric neighborhood, LDOF, LOF techniques, and recently introduced RDOS.