A distance-based method for outlier detection on high dimensional datasets

Abstract

Tasks such as classification, clustering, and regression require the identification and elimination of outliers as part of the preprocessing of data. Without an adequate processing of outliers the results of data analysis will be biased and inexact. This article proposes a distance-based method for the detection of outliers in multivariate datasets. The proposed method takes advantage of the principal components for avoiding problems of collinearity in datasets and the high concentration of variance is used to increase the separation between outliers and inliers. The proposed method was compared against four outlier detection methods from the literature, two deterministic and two stochastic. The datasets used in the comparison were generated according to different works in the literature. These datasets follows different distributions and contains different amount of outliers and inliers, and different number of variables and instances. Per each distribution, low and high dimensionality were considered. Unlike to other methods in the state-of-the-art, the proposed method does not require apriori the definition of any value for its operation. Also, the calculation of distances of elements in the dataset takes lower processing time. According to the experiments, the proposed method is suitable to deal with dataset contaminated with low and high proportions of outliers, low and high dimensions, and simmetric and assimetric distributions. Also supports colineality in data.