Database Reference
In-Depth Information
where x
i
represents the data points in the attribute space and
y
i
the target attribute.
Assume now that these data have been obtained by sampling of an unknown
function
f
which belongs to some function space
V
defined over
d
. The sampling
process was disturbed by noise. The aim is now to recover the function
f
from the
given data as faithfully as possible. We distinguish between
classification
, where
the target values
y
i
are from a discrete set of classes, e.g., from {
1, +1} for binary
classification, and
regression
where
y
i
are from a continuous spectrum. In what
follows we mainly focus on classification having in mind that sparse grids can be
used for regression, too [Gar06, Gar11]. In classification the function
f
is also called
classifier
.
Scoring is increasingly used for personalization and may also be applied to
recommendations. An advantage of scoring is that we can include many attributes
characterizing the user behavior in x
i
. This may be user-centric attributes like age
and gender, transactional attributes like number of clicks or revenue, and many
other attribute types like time, channel, or even weather. The disadvantage of
scoring is the limited number of single attribute values it can handle in general.
This renders a
direct
application of scoring for recommendations of many products
virtually impossible.
There are different approaches to scoring-based recommendations. The most
simple is to use the recommendations as target attribute, i.e., each recommended
product corresponds to a target class. A more sophisticated approach is to use the
success of the session (revenue or in case of classification indicator of orders in the
session) as target attribute and the recommendation as a special set of
control
attributes. Thus, in each recommendation step, we select the control attributes to
maximize
f
(x). (Note that depending on the function class of the classifier
f
, this
may result in a complex optimization problem. But this is not the main task of
scoring and hence will not be considered here.)
R
Example 7.1
Consider a small web shop. Suppose we need to select one of the
three on-site banners at each category and product page. Therefore, the banners
represent the control attribute and thus the recommendations. We further assume
that in each step of the session (product or category page view), a user is charac-
terized by four attributes: age, gender, number of clicks in current session, and how
many products are already in her/his basket. The target attribute is 0 if no order was
placed within the session and 1 if something was ordered.
Table
7.1
shows three sample sessions. In the first step of session
A,
the user is
considered to be unknown and hence his/her user-specific attributes age and gender
have missing values (represented by character “?”). In the first step, the banner
b1
was recommended to his/her. We know from history that he/she has bought nothing
in this session, so the target attribute is 0 in all steps of the session. The second step
is very similar to the first one except that banner
b3
was recommended. In the third
step, he/she added a product to his/her basket. In the fourth step, he/she signed in to
the shop and now his/her age and gender are considered to be known.
Session
B
represents a registered user, who was already recognized at the
beginning of the session, e.g., by a cookie. This user finally placed an order.
