Some Practical Aspects on Multidimensional Scaling of Compositional Data 1 Some Practical Aspects on Multidimensional Scaling of Compositional Data

To visualize the data with Multidimensional Scaling methods we approximate a given dissimilarity matrix {matrix of di erences among observations{ to obtain a con guration of points in low (two) dimensional real (usually) Euclidean space. The Multidimensional Scaling methods input is a dissimilarity matrix and to construct such a matrix a suitable measure of di erence between observations is needed. In our work we discuss applications of di erent dissimilarity measures, relations between them and their (un)suitability in case of compositional data. We present results of Multidimensional Scaling methods applied to real compositional data sets to visualize all these relations. Visualizations also con rm our theoretical results and show which dissimilarity measures are coherent with the compositional nature of the data.