where Z is a I ×2 matrix representing the objects in 2D Euclidean space and the
matrix Yk is the 2 dimensional representations of the category values of the attribute
k. Both Z, the coordinates of objects, and all Yk’s, the coordinates of the
attribute category values, can be plotted in a joint space which is called a biplot
[15]. Essentially, Z−GkYk gives the differences (or error) between the position of
the individual products and the positions of the category centroids they belong to
for variable k. Ideally, no error would exist and all products in the same category
would be fully homogeneous and coincide with the position of their category. As
there are more attributes and the products fill in different categories on the different
attributes, (17.8) simply measures the squared distances of the products relative to
their category centroids, hence how homogeneous the products are. The matrix Mk
removes the contribution to the error for any product having a missing value for attribute
k. Equation (17.8) can be minimized using an alternating least squares (ALS)
procedure called Homals [13, 34].
From Homals, it is an easy change to NL-PCA by imposing extra restrictions
on the category points. Numerical and ordinal attributes can be incorporated in the
Homals framework when we impose a rank-1 restriction of the form