A Deterministic Multidimensional Scaling Algorithm for Data Visualisation

In this paper, we present I-PACK, a deterministic layout algorithm for embedding a data set X in 2D provided that distances (delta_uv )_u,visinX, between data items are given or can be computed. The layout reflects well similarities and dissimilarities between items and it is computed in quasi-linear time. Experimental comparisons with other multidimensional scaling algorithms show that: i) our algorithm has similar performance when the aspect ratio A = max_u,v)(delta_uv)/ min_u,v)(delta_uv) is small (i.e. log₂A < 10) and ii) the larger the aspect ratio, the better I-PACK performs with respect to other MDS algorithms. This is also true when data can be "naturally" clustered