ClustanMDS Limitations ClustanMDS requires a proximity matrix of type lower diagonal, symmetric. The similarity between any case and itself, which corresponds to the diagonal elements, is considered to be maximum and equal to the similarity between two identical cases – therefore diagonal elements are not required. However, if you input a square or symmetric proximity matrix with the diagonal present, the diagonal elements will be assumed maximum. ClustanMDS does not provide TORSCA, arbitrary or input starting configurations of KYST – from tests with proximity matrices containing definite structure we found that trials using random starting configurations usually converge to a lowstress solution in most instances. With 100 trials of the Mammals Milk data (25 cases in 5 dimensions), local minima of around 6% stress were found in 18 trials, half of which failed to converge in 5000 iterations; the other 82 trials all converged to an excellent twodimensional fit of less than 1.5% stress. These and other tests indicate that a large number of trials with random starts are as good as, or better than, preprocessing configuration options. ClustanMDS does not use the general Minkowski R metric distance function provided in KYST. We have adopted the KYST default of R=2, which corresponds to Euclidean distance which we believe should satisfy the majority of users. This change alone substantially improved the program code performance. ClustanMDS cannot split by rows or groups, as provided in KYST for unfolding analysis, but operates on a standard proximity matrix as read or computed in ClustanGraphics.
