ABSTRACT
In this chapter we will introduce and motivate robust unsupervised sample reduction methods which are not based on probabilistic assumptions on the data generating distribution. We will start from one of the first attempts to robustify sample reduction, namely, partitioning around medoids. We then will proceed with trimmed k-means. Finally, we review snipped k-means, a generalization of trimmed k-means tailored for component-wise outliers. The reader is also referred to Banerjee and Dave (2012) for a review of some topics which are not covered here, including robust hierarchical clustering and robust fuzzy clustering.
