RDPOD: an unsupervised approach for outlier detection

Outliers are the data points which deviate significantly from the majority of the data points. Finding outliers is an important task in various applications, especially in data mining. The unsupervised technique is very popular to mine outliers in a dataset over supervised techniques. Various unsupervised approaches have been proposed over the last decades. Clustering-based, distance-based, and density-based outlier approaches are found to be successful for detecting outlier points. However, the main focus of clustering-based method is to identifying clustering structure. Many distance-based and density-based techniques are not suitable for varying density datasets, and they are also very sensitive with their parameter (number of nearest-neighbor (k)). In this paper, we propose a hybrid approach named RDPOD, which utilizes distance-based and density-based clustering approaches efficiently for identifying the density of each point correctly. We obtain local density and relative distance of each data instance. From this density and distance information, we identify outlier points. Experimental results with real-world datasets show that our proposed approach outperforms the popular techniques LOF, LDOF, symmetric neighborhood, and recently introduced approaches NOF and RDOS.

Automated summary

The paper proposes a hybrid approach named RDPOD, which efficiently utilizes distance-based and density-based clustering methods to correctly identify the density of each point. Analysis of experimental results shows that our proposed approach outperforms other popular techniques in detecting outlier points.