Robust multidimensional scaling

Modern technology enables the collection of vast quantities of data. Smart automatic data selection algorithms are needed to discover important data structures that are obscured by other structure or random noise. We suggest an efficient and flexible algorithm that chooses the "best" subsample from a given dataset. We avoid the combinatorial search over all possible subsamples and efficiently find the datapoints that describe the primary structure of the data. Although the algorithm can be used in many analysis scenarios, this paper explores the application of the method to problems in multidimensional scaling.