Personalized Fashion Advice - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 223

Fig. 8.8 User model for CBR: Exemplification of the user model

• Jeans

• Shoes

• Basics

For these different categories, we model the user preferences separately. Each
category consists of attributes describing color preferences for the category, the brand,
the material, or the price. For these attributes, we model what a user likes and what
not. This allows us to look for color-matching clothes with colors the user prefers
in one category but does not like for another one. We also model the preference for
certain item attributes such as cut or imprints. Figure 8.8 depicts the structure of the
user model.

8.5.2 CBR Approach

The model described in the previous section constitutes the basis for our personalized
CBR approach. Given the model shown in Fig. 8.8 , our CBR recommender learns
from past purchases of the user what features, for what category, lead to a purchase,
and which do not. This is done per user. Our CBR takes as an input all available
purchases of the user, which consist of an article description with all features, the
price and the information if the item was bought or returned. We have additional
information about the user, containing demographic information and body size data.
We also have information about the price range a user is willing to pay for an item
for certain categories. One user may be willing to pay

e 50 for a shirt while others

will abandon shirts of more than 30. These information are handled as explicit

constraints by the user. Implicit constraints are derived from the article features
that are bought or returned. Therefore, the CBR consist of two types of constraints—
explicit ones set by the user and implicit ones derived fromuser purchases. Both types
of constraints are handled differently. If the CBR detects a violation of an explicit

e