17.3.1 Multidimensional Scaling
The first approach to create a product catalog map that we describe is based on MDS.
As was mentioned in Section 17.2.3, the basis of an MDS map is a dissimilarity
matrix. To compute a dissimilarity matrix Δ from the product catalog, we need a
dissimilarity measure. This measure should be able to cope with the specific data
contained in a product catalog, that is, it should be able to handle missing values
and numerical, categorical, and multi-valued categorical attributes.
Many popular (dis)similarity measures such as the Euclidean distance, Pearson’s
correlation coefficient, and the Jaccard similarity measure are not able to handle
all of these attribute types. Moreover, they can not handle missing values naturally.
Therefore, we use a dissimilarity measure which is an adaptation of the general
coefficient of similarity proposed by Gower [14] and was introduced in [23]. Note
that the MDS based product catalog map approach can also be used with other
dissimilarity measures, such as co-purchases or item-item correlations.