17.4 Determining AttributeWeights using Clickstream Analysis
Both product catalog map approaches introduced in the previous section consider
all attributes of a product as equally important to the user. However, we showed that
attribute weights can be incorporated in the dissimilarity measure used for MDS and
also NL-PCA can be adapted to incorporate different weights for attributes. This
holds for most visualization methods, including self-organizing maps and treemaps
which were described in Section 17.2.
In this section, we will introduce an approach to determine these weights automatically
using clickstream data. For every product, we count how often it was sold
during some period. In our application, shown in Section 17.6, we actually counted
outclicks, which are clicks out of the site (we used data of a price comparison site)
to a shop where the product can be bought. For ease of readability, we use the term
sales instead of outclicks during the remainder of this paper. Using these counts and
the product attributes, we estimate a Poisson regression model, which is a model
belonging to the class of generalized linear models. Using the coefficients of this
model and their corresponding standard errors, we compute t-values which form
the basis of our attribute weights. This approach was described earlier in [22].