Also, weights wi j can be specified to force some dissimilarities to be fit better than
others.
The dissimilarity measure we use is able to handle missing values. However, dissimilarities
based on only a couple of nonmissing (and maybe even unimportant)
attributes are more unreliable than dissimilarities for which no dissimilarity scores
were missing. Therefore, the latter should receive higher weights. This can be done
by defining the weights to be used in (17.5) as the weighted proportion of nonmissing
attributes used for pair i j