Above the product catalog map, ther

Above the product catalog map, there is a list with all product attributes. Each attribute
can be selected by clicking on its check box. If one selects a certain attribute,
the category points of this attribute are added to the map. The category points of the
selected attributes not only help the user to interpret the map, they are also tools to
navigate through the map. By clicking on a category point on the map this category
is added to the selection list.
The selection list, which is shown to the right of the list of attributes, determines
the products that are highlighted on the map. The set of highlighted products is determined as follows. For each of the selected attributes, a shown product should
belong to at least one of the selected categories.
When the selection list has been adapted by adding or removing a category point,
a couple of things are modified in the visualization of the map. The first thing is that
all products satisfying the new constraints are colored red, while all other products
are colored blue. Furthermore, a number of products is randomly selected from this
set to be highlighted.
Since a selection will often lead to a subspace of the map, it is also possible to
zoom in on this part of the map. However, there is no guarantee that all points in this
subspace satisfy all constraints imposed by the selection list. We have chosen not
to delete these product points, since these may be interesting to the user. Although
these products do not satisfy all the demands of the user, they are very similar to the
products that do and may have some appealing characteristics the user until then did
not think of.
In both the complete and the zoomed in map, the user can click on the highlighted
products to get a full product description of this selected product, which is shown
at the right side of the application. However, by moving over both a full color or
a monochrome image, a tooltip is shown containing an image of the product and
the values of some of its most important attributes. Furthermore, the values for the
attributes in the selection list are shown in this tooltip, colored green when they
match the preferred category value and red when they do not. Since the GUI is
based on a single NL-PCA map, this map too can be computed offline just as the
MDS product catalog map.
Since the quality of NL-PCA maps may become quite poor when having a lot of
missing values, we removed attributes having more than 50% missing values. Then,
we also removed products having more than 50% missing values on this limited set
of attributes. This resulted in a set of 189 MP3-players described by 19 attributes,
namely the attributes shown in Table 17.2.
In the NL-PCA algorithm, we will treat all numerical attributes as being ordinal,
because of two reasons. In the first place, many of the numerical attributes do not
have a linear interpretation for users, such as, for example, the memory size. The
second advantage of using the ordinal transformation is that due to the underlying
monotone regression procedure some adjacent categories can be merged into a single
category point. Since a numerical attribute has a lot of categories (i.e. all unique
values in the data set), visualizing all these categories may become unclear and
selection using these category values may become useless, since a lot of category
values only belong to a single object. Using an ordinal transformation this becomes
less of a problem, since categories with a small number of objects are often merged
with their neighbors.
In Figure 17.10, two biplots are shown resulting from the NL-PCA algorithm. A
biplot visualizes both the products labeled by their attribute value and the category
points also labeled by their value. Also, the origin is plotted in the biplots. By the
design of the NL-PCA method, ordinary products (products having attribute values
that are similar to many other products) should be close to the origin, while more
distinct products should be far away. This also holds for the categories. Since both biplots in Figure 17.10 are plots of numerical attributes, the category points are on
a line. Like in the MDS map, also here both Price and Memory Size correlate with
each other and are well represented, in this case on the second dimension.

Results (English) 1:

Above the product catalog map, there is a list with all product attributes. Each attributecan be selected by clicking on its check box. If one selects a certain attribute,the category points of this attribute are added to the map. The category points of theselected attributes not only help the user to interpret the map, they are also tools tonavigate through the map. By clicking on a category point on the map this categoryis added to the selection list.The selection list, which is shown to the right of the list of attributes, determinesthe products that are highlighted on the map. The set of highlighted products is determined as follows. For each of the selected attributes, a shown product shouldbelong to at least one of the selected categories.When the selection list has been adapted by adding or removing a category point,a couple of things are modified in the visualization of the map. The first thing is thatall products satisfying the new constraints are colored red, while all other productsare colored blue. Furthermore, a number of products is randomly selected from thisset to be highlighted.Since a selection will often lead to a subspace of the map, it is also possible tozoom in on this part of the map. However, there is no guarantee that all points in thissubspace satisfy all constraints imposed by the selection list. We have chosen notto delete these product points, since these may be interesting to the user. Althoughthese products do not satisfy all the demands of the user, they are very similar to theproducts that do and may have some appealing characteristics the user until then didnot think of.In both the complete and the zoomed in map, the user can click on the highlightedproducts to get a full product description of this selected product, which is shownat the right side of the application. However, by moving over both a full color ora monochrome image, a tooltip is shown containing an image of the product andthe values of some of its most important attributes. Furthermore, the values for theattributes in the selection list are shown in this tooltip, colored green when theymatch the preferred category value and red when they do not. Since the GUI isbased on a single NL-PCA map, this map to can be computed offline just as theMDS product catalog map.Since the quality of NL-PCA maps may become quite poor when having a lot ofmissing values, we removed attributes having more than 50% missing values. Then,we also removed products having more than 50% missing values on this limited setof attributes. This resulted in a set of 189 MP3-players described by 19 attributes,namely the attributes shown in Table 17.2.In the NL-PCA algorithm, we will treat all numerical attributes as being ordinal,because of two reasons. In the first place, many of the numerical attributes do nothave a linear interpretation for users, such as, for example, the memory size. Thesecond advantage of using the ordinal transformation is that due to the underlyingmonotone regression procedure some adjacent categories can be merged into a singlecategory point. Since a numerical attribute has a lot of categories (i.e. all uniquevalues in the data set), visualizing all these categories may become unclear andselection using these category values may become useless, since a lot of categoryvalues only belong to a single object. Using an ordinal transformation this becomesless of a problem, since categories with a small number of objects are often mergedwith their neighbors.In Figure 17.10, two biplots are shown resulting from the NL-PCA algorithm. Abiplot visualizes both the products labelled by their attribute value and the categorypoints also labeled by their value. Also, the origin is plotted in the biplots. By thedesign of the NL-PCA method, ordinary products (products having attribute valuesthat are similar to many other products) should be close to the origin, while moredistinct products should be far away. This also holds for the categories. Since both biplots in Figure 17.10 are plots of numerical attributes, the category points are ona line. Like in the MDS map, also here both Price and Memory Size correlate witheach other and are well represented, in this case on the second dimension.