We minimize σr(Z) using the SMACOF algorithm [10] which is based on majorization.
One of the advantages of this method is that it is reasonable fast and
that the iterations yield monotonically improved Stress values and the difference
between subsequent coordinate matrices Z converges to zero [10]. This property
has an important and vital consequence for dynamic visualizations: the algorithm
produces smooth changes to the points in the display leading to a (local) minimum
solution of (17.5). In effect, the objects follow a smooth trajectory on the screen.
The resulting maps may look overwhelming to the user when the number of
products is large. To make the map more appealing to the user, a small number of
products is highlighted by showing larger sized and full color images, while other
products are represented by a smaller monochrome image. These highlighted products
are helpful to the user to get a quick overview of the map. Therefore, it is nice
when these products represent different groups of products in this map. This was
done, by first clustering the products in the map using k-means clustering [17]. We
decided to perform a k-means clustering on the map Z instead of a hierarchical clustering
procedure on the original dissimilarities for two reasons. First, this procedure
is faster and, second, it is consistent with the visualization, that is, there is no overlap
between the clusters in the map. In each cluster, one product is chosen to be highlighted,
that is, the product closest to the cluster center based on Euclidean distance.
In Section 17.6, we show a prototype of this approach using a product catalog of
MP3-players.