An Outlier Detection Algorithm based on KNN-kernel Density Estimation

Abstract

The importance of outlier detection is growing significantly in a various fields, such as military surveillance,tax fraud detection, telecommunications, terrorist activities, medical and commercial sectors. Focusing on this has resulted in the growth of several outlier detection algorithms, mostly based on distance or density strategies. But for each approach, there are inherent weaknesses. The distance-based techniques have a local density issue, while the density-based method has a low-density pattern issue. In this article, we present an unsupervised density-based outlier detection algorithm to address these shortcomings. In the proposed approach, each object is assigned a local outlying degree, which indicates how much one point in its locality deviates from the other. The local outlying degree focuses explicitly on the concept of local density, which is defined as a relative measure of the local density of the object to the local density of its neighbour. The proposed approach uses a measure of k nearest neighbour kernel density (NKD) to estimate the density. Besides, our proposed algorithm used three different categories of nearest neighbours, k nearest neighbour (kNN), reverse nearest neighbour (RNN), and shared nearest neighbour (SNN) to make our systems more flexible in modeling different local data patterns. Formal analysis and extensive experiments on artificial and UCI machine learning repository datasets show that this technique can achieve better outlier detection performance.