Information Technology Reference
In-Depth Information
constraint, the article will be automatically rejected. In our case, currently only price
constraints are set and used. In case an article is not falling into the user-defined price
range, this article will not be recommended to the user, even if the article matches the
implicit constraints. Our CBR approach consists of two phases. The learning phase,
which builds a personalized user model, and the recommendation phase, where the
user model is compared to an article to compute the similarity which is the basis for
the recommendation.
The learning phase takes (procedure depicted inAlgorithm1) into account all users
U ={
u
1
,
u
2
,...,
u
n
}
, purchased articles
A ={
a
1
,
a
2
,...,
a
n
}
and if they bought
or returned it
I
(
u
,
a
)
={
return
,
buy
}
. An article itself consists of a set of features
F
defined as
a
. A user model consists of weights for
all features of all articles the user interacted with: um
={
f
1
,
f
2
,...,
f
n
}
where
f
∈ F
.
The code example Algorithm 1 shows the general computation of these implicit
constraints. Explicit constraints, as said, are not computed but a predefined set, which
is filled in the learning phase with information from the given user data.
The goal of the learning phase is to have a user model representing an ideal article
for the user, which can be compared to other articles. The model consists of different
sub-models for the different categories.
The recommendation phase compares a given article
a
using a similarity function
sim
={{
f
1
,w
}
,
{
f
2
,w
}
,
{
f
n
,w
}}
, see Algorithm 2. The
similarity function sim compares the user model with the given article. The function
also takes thresholds defining when an article is marked as recommendable or not.
(
a
,
thresholdPositive
,
thresholdNegative
)
Algorithm 1
Learning constraints for the User Model.
for all
i
∈
I
do
buy
then
for all
f
∈
a
do
Add
f
to user model and increase weight
end for
else
for all
f
∈
a
do
Add
f
to user model and decrease weight
end for
end if
end for
if
i
=
If an article matches the user model, it will be recommended. Similarity is cur-
rently measured by counting the number of features overlapping the user model and
the article. Beside articles marked as recommendable, the CBR will also mark arti-
cles as definitely not recommendable (e.g. by failing the price constraint). As for the
following evaluation, we focus on correct recommendations of articles.
Search WWH ::
Custom Search