A Novel Approach in Multivariate Outlier Detection

Abstract

Detecting outliers in the multidimensional space is as important as detecting them in a single dimension. The term "outlier" refers to the observation which is in some way inconsistent with the rest of the observations in a data set. Outliers can lead to incorrect calculation of sample parameters, and thus to poor estimation of population parameters. Definitions of outliers are numerous. The most commonly cited definition is that it is "an observation that deviates so much from other observations as to arouse suspicion that it was generated by a different mechanism. Multivariate outliers are most commonly detected using the Mahalanobis distance. In this research, the statistical I-distance method is thoroughly explained, applied and compared with Mahalanobis distance, since they have similar nature and calculation process. I-distance method, as a metric in an n-dimensional space, has been originally devised in order to rank countries according to their level of development, based on several indicators. There are many improvements оf this method that led to its widespread use, such as multivariate outlier detection. This research is conducted on 30 point guards in the NBA league and the values of nine indicators were measured in order to detect players with specific set of skills.