Extended isolation forest – application to outlier detection in geomagnetic data

Abstract

The aim of this study is to present a method for detection of outliers in the time series of total intensity of geomagnetic field using Extended Isolation Forest algorithm. The method is consisted of three steps: 1) generation of additional features that take into account the regular daily variation and smooth behaviour of normal data, 2) detection of potential outliers based on ensemble of extended isolating trees and 3) subsequent refinement based on difference between the outlier and its replacement with interpolated value. Application of the method for detection of outliers in yearly time series of the total geomagnetic field at Ak-Suu and Kegety stations showed that the algorithm identifies both global and contextual outliers. Average classification metrics for the method are characterized as high and have the following values: precision 94.3%, recall 93.9% and F-score 94.5%, and probabilities of errors of the first and second kind are comparable to similar algorithms used for detection of outliers in magnetograms of different sampling rate.