Database Reference
In-Depth Information
4.2 Background
After the pioneering work done in the Grouplens project in 1994 [
5
], collaborative
filtering (CF) soon became one of the most popular algorithms in recommender
systems. Many variations of this algorithm have also been proposed such as hybrid
approaches of combining CF with content-based filtering [
7
,
17
-
19
] or adopting
different weighting schemes [
6
,
20
]. In this chapter, we will use the traditional CF
proposed in the Grouplens project as one of the comparison methods. Therefore, the
remainder of this section will focus on this algorithm.
The assumption of CF is that people who agreed in the past tend to agree again in
the future. Therefore, CF first finds users with tastes similar to those of the target
users. CF will then make recommendations to the target user by predicting the
target user's rating of the target item based on the ratings of his/her top-
K
similar
users. User ratings are often represented by discrete values within a certain range,
for example, 1-5. Here 1 indicates an extreme dislike of the target item, while
5 shows high praise. Let
R
UI
be the rating of the target user
U
on the target item
I
.
Thus,
R
UI
is estimated as the weighted sum of the votes of similar users as follows.
R
UI
¼
Z
X
V
R
U
þ
w
ð
U
;
V
Þð
R
VI
R
V
Þ
;
(4.1)
2
C
where
R
U
and
R
V
represent the average ratings of the target user
U
and every user
V
in
U
's neighborhood,
, which consists of the top-
K
similar users of
U
.
w
(
U
,
V
)
is the weight between users
U
and
V
, and
z
C
P
V
wU
;
V
1
¼
is a normalizing constant to
ð
Þ
normalize total weight to one. Specifically,
w
(
U, V
) can be defined using the
Pearson correlation coefficient
[
5
].
P
I
ð
R
V
Þ
P
I
ð
R
UI
R
U
Þð
R
VI
w
ð
U
V
Þ¼
q
(4.2)
;
;
P
I
ð
2
2
R
UI
R
U
Þ
R
VI
R
V
Þ
where the summations over
I
are over the common items for which both user
U
and
V
have voted.
As we can see, the traditional CF models user-to-user relations based purely on
user rating similarities and does not utilize at all the semantic friend relations among
users. However, such semantics are essential to the buying decisions of users. In the
following sections, we are going to present a new paradigm of recommender systems
which improves the performance of recommender systems by using the semantic
information in social networks.
4.3 Yelp.com
For this research, we collect a dataset from a real online social network Yelp.com.
As one of the most popular Web 2.0 websites, Yelp provides users with local
searches for restaurants, shopping, spas, nightlife, etc. Besides maintaining the
Search WWH ::
Custom Search