Most classification methods assume that the objects to be classified have been embedded in a Euclidean space. If another measure of dissimilarity is preferred, then the objects can be embedded in a Euclidean space via multidimensional scaling (MDS). The use of MDS to preprocess data for classification poses various conceptual and technical challenges. Traditional MDS is unsupervised, more commonly used for clustering than for classification. Because traditional MDS maps a finite set of objects into Euclidean space, it is not clear how to construct a classifier that can be applied to new cases. Novel formulations of MDS address these challenges.