![](https://csdnimg.cn/release/download_crawler_static/15633586/bg4.jpg)
4 Deqing Yang, Yanghua Xiao, Hanghang Tong, Junjun Zhang, and Wei Wang
The rest of this paper is organized as follows. We first display our empirical results
in Section 2 which are the basis of our recommendation algorithm. In Section 3, we
elaborate the detailed procedure of tag recommendation algorithm. In Section 4, we
present our experiments for evaluating algorithm performance. We survey the related
works in Section 5 and conclude our paper in Section 6.
2 Empirical Study
In this section, we conduct empirical studies on the collective tagging behaviors of
Weibo users. The empirical findings construct the basis of our tag recommendation
algorithm. We first introduce our dataset.
Dataset: We first randomly selected 3,000 Weibo users as seeds, then crawled their fol-
lowers and followees. Thus there are more than 2.1 million users and 875,186 unique
user tags in total. Besides tags, the following relationships between the users were
fetched. All data were crawled before Oct. 2013. The statistics show that only 55.01%
of these users, i.e., about 1.15 million users have at least one tag.
2.1 Homophily in Tagging Behavior
Homophily is a tendency that an interaction between similar people occurs with a higher
probability than among dissimilar people [19]. Homoplily was shown to be a univer-
sal phenomenon across a variety of social media platforms such as Twitter [28]. More
specifically, the Twitter users following reciprocally (mutual fans) tend to share topi-
cal interests, have similar geographic and popularity [15]. Thus, an interesting question
arises: do close social relationships in Weibo also imply similar profiles or tags? To
answer this question, we first distinguish three important types of social relationships
among Weibo users: following (follower), followed (followee) and following recipro-
cally (mutual fan)
4
. Next, we will empirically study the effects of these three relation-
ships on tag similarity. At first, we define two types of tags for a Weibo user u.
Definition 1 (Real Tags). If u originally labels him/erself with some tags, these tags
are referred to as u’s real tags and denoted by RT
u
.
Definition 2 (Collective Tags). The tags that are most frequently used by u’s friends
are referred to as u’s collective tags and denoted by CT
u
.
To find the tags in CT
u
, we define a score function tf(t) to quantify the likelihood that
tag t belongs to CT
u
. The tf(t) function is defined as
tf(t) =
r(t)
t
′
∈T (N eg(u))
r (t
′
)
(1)
where N eg(u) is u’s friend group and r(t) is the number of users in Neg(u) who have
used tag t. T (Neg(u)) represents the tag set used by the users in N eg(u). We denote
the score function as tf because it is equivalent to the term frequency in document
retrieval. The larger the tf(t) is, the more likely the tag t belongs to CT
u
. If |CT
u
| is
limited to k, we select the top-k tags from T (Neg(u)) according to tf(t) value. In the
following text, we refer to tf(t) as the frequency based tag ranking score.
4
In this paper, we often refer to these three social relationships in Weibo as friend.