Personalized Fashion Advice - Smart Information Systems: Computational Intelligence for Real-Life Applications - page 228

80

70

60

50

40

30

20

10

0

Fig. 8.10 Results for the accuracy measure of the different approaches

value. In the next section, we present a short analysis of the different attributes and
the information gain they offer.

8.6.1 Information Value of Attributes

We noted that the recommendation quality depends on the way our algorithm treats
available attributes. We have seen that treating all features equally negatively affects
the performance. We suppose this to be due to individual features carrying more or
less valuable information about the relation between users and items. We determined
each feature's value in terms of “mutual information” with the information about
whether an article had been returned or not. Hereby, the system encodes the target
quantity as a binary variable. Customers may either return or keep an article. On the
other hand, features' domains include categorical, ordinal, and numeric value ranges.

p ( x , y )

I X , Y = p ( x , y ) log 2 (8.3)

p ( x ) p ( y )

x ∈ X , y ∈ Y

Equation 8.3 illustrates howwe compute mutual information [ 36 ]. X and Y represent
two random variables. In our case, Y refers to the customer keeping or returning an
article. X represents the feature for which we seek to determine the information
contained. Note that we employ the logarithm with base 2 to obtain information in
term of bits. Alternatively, we could use logarithms to base e which would provide
information in terms of nits.
Figure 8.11 depicts how the mutual information distributes across different types
of features.We notice that only feedback-related features carry a considerable amount
of information. Unfortunately, feedback-related features exhibit highest levels of
sparsity. In addition, customer who had not bought any article yet will lack this