Exemplar Sampling

Abstract

An exemplar sample is a small subset of a dataset that forms a characterization of the dataset and can also be used as a list of targets for further inspection. Traditionally exemplars are encountered in cluster analysis techniques such as k-medoids, where each exemplar represents a different cluster. This article extends the concept of exemplar sampling by constructing four criteria for choosing a sample: similarity, distinctness, exhaustiveness, and typicalness. Each criterion illustrates a different way of characterizing the dataset. We describe an adapted version of the CLARANS algorithm which can be used to efficiently draw exemplar samples, and show some examples on test datasets. Exemplar sampling is proposed for use in finite population inference, as an alternative to methods such as stratification or balancing.