Pseudoinverse in clustering problems (Q2480500)

The paper deals with the application of pseudoinverse methods, especially, with their development as the pseudoinverse-perturbation theory in solving clustering problems. The solution algorithm for a clustering problem is recurrent. At the first step, the whole set of vectors is arbitrarily divided into two groups. During the subsequent steps, each group is sequentially ``improved'' by choosing and eliminating elements that are not peculiar to the group to which they belong. In the basic case, the authors propose to compute the degree of correspondence based on different distances of each group. A choice of such distances determines the clustering technique. In particular, the basic case suggests considering distances to hyperplanes generated by vectors of corresponding cluster groups. The authors consider various cases of clustering where the clusters are corresponding hyperplanes. The pseudoinverse technique allows generating constructive explicit formulas for correspondence distances, which provides efficient computational implementation of the corresponding algorithms.