A comparison of outlier detection methods: exemplified with an environmental geochemical dataset

Abstract

Three outlier detection methods of range, principle component analysis (PCA), and autoassociation neural network (AutoNN) approaches are introduced and applied to an environmental geochemical dataset in Sweden. Each method uses a different criterion for the definition of outlier. In the range method, the number of outlying values of one sample is determined as the outlying sample measurement parameter. The distance of sample scores in the principal components from the coordinate origin is suggested as the parameter for the PCA method. The total sum of error squares between the measured and predicted values is proposed as the parameter for the AutoNN approach. The results of the three methods are comparable, but differences exist. A combination of all the methods is recommended for the development of a better outlier identifier, and further analyses on the detected outliers should be carried out by integrating geological and environmental information.