GAP method: A dichotomization technique for outlier detection and its application to wildlife GPS data

Abstract

Abstract Outliers in datasets are a concern for analysts as disturbances or signals, and various techniques for their detection have been proposed. Some of them separate outliers as output, but others do not. The latter provide measures of how likely each data point is an outlier, but they leave the threshold-setting for separating outliers as a task for analysts. We developed a technique to help analysts perform this task. This technique uses value gaps between adjacent data pairs in a univariate dataset, where the data are sorted in ascending order of value. Its core process is to find the largest gap in the upper range of the dataset and remove the data above the gap as outliers; its supplementary process is to repeat the core process for the dataset after removal. Analysts need to decide when to stop this iteration. However, this process leaves analysts with only a few reasonable options for the decision. This method applies to any dataset, such as a time series or multivariate dataset, if a ratio-scale measure for quantifying the degree of data being an outlier is given. We demonstrate how to implement this technique using wildlife GPS data and discuss the uniqueness and usefulness of the approach.