Database Reference
In-Depth Information
Table 7.1 Example of training set of three banner recommendations for classification
ID attribute
Input attributes
Control attribute
Target attribute
Session
Age
Gender
Clicks
Basket
Banner
Ordered
A
?
?
1
0
b1
0
A
?
?
2
0
b3
0
A
?
?
3
1
b2
0
A
24
f
4
1
b2
0
B
19
m
1
0
b3
1
B
19
m
2
0
b2
1
B
19
m
3
1
b2
1
B
19
m
4
2
b3
1
C
?
?
1
0
b2
0
The third session
C
represents an unknown user who clicked only once and then
left the shop.
Based on such historic data, by means of a classification technique, we now can
construct a classifier
f
that assigns a target value
y
to each attribute vector x
5
(the four input attributes and the control attribute), i.e.,
y ¼ f
(x). The higher the
y
,
the higher is the probability that an order will be placed inside the session. Since
the input attributes are fixed, we maximize
f
in each step with respect to the
control attribute.
To illustrate this procedure, consider a session step of a 56-year-old woman who
has already done 3 clicks and added two products to the basket, i.e., x
¼
(56,
f
,3,2,
x
c
), where
x
c
represents the value of the control attribute. Let
f
(x) be 0.6 for
x
c
¼ b
1
,
0.34 for
x
c
¼ b
2
, and 0.85 for
x
c
¼ b
3
. Then we would select banner
b
3
as
recommendation.
Thus, in each step of the session, we use the current values of the input attributes
and find the optimal value of the control attribute. The corresponding banner is
recommended.
The construction of the classifier, i.e., the learning, is performed either offline on
historic data stored in the form of Table
7.1
or online, after each session terminated
(e.g., by a timeout mechanism).
∈
R
■
Although Example 7.1 is very simple, it reveals the power of the scoring
approach for recommendations. Unlike basket analysis or collaborative filtering
(which will be studied in the next chapter), it considers the recommendation task as
control problem taking into account the effect of recommendations. As mentioned
before, it also allows to include many different attributes into prediction. Scoring is
used to calculate banner recommendations, special offers during the checkout
process, and even for personalized navigation. Of course, it also bears some
disadvantages: the control-theoretic approach is very limited and so is the ability
to handle large numbers of recommendation items.
There exist many algorithms for classification and regression. Widely used
approaches are nearest neighbor methods, decision tree induction, rule learning,
and memory-based reasoning. There are also classification methods based on
