Information Technology Reference
In-Depth Information
d
∈
R
where
v
is the video feature vector by aggregating title, tag, and description,
d
is the feature vector for the
k
th topic by remaining the top probable words,
and
d
is the dimension of the vocabulary.
z
k
∈
R
5.3.2.2 Long-Term User Modeling
For long-term user modeling on YouTube, we construct user profiles by aggregating
the registration information (“AboutMe”) and the social activity data, recorded as
a feature vector
u
d
in the VSM. The personalized video relevance based on
long-term interest is accordingly calculated as:
∈
R
v
T
u
√
v
T
v
√
u
T
u
s
long
(
v
,
u
)
=
(5.2)
We then obtain the final relevance score by integrating the long-term and short-term
interest as a linear combination:
s
(
v
,
u
)
=
ʻ
t
s
long
(
v
,
u
)
+
(
1
−
ʻ
t
)
s
short
(
v
,
u
)
(5.3)
where
ʻ
t
is the weighting parameter selected on a daily basis. The logic behind
Eq. (
5.3
) can be interpreted by the following example. A user is discovered to follow
the trending topic “Euro Cup 2012” on Twitter, which indicates the short-term inter-
est. By referring to the YouTube activity history or profile, we observe that he/she
is a Beckham fan or comes from England, which indicates the long-term interest.
Combining both, we are confident to recommend him/her the England match videos
or Beckham news videos related to Euro Cup 2012.
In order to evaluate the performance of the proposed recommender, we employ the
videos that favored, rated or commented by the overlapped users as the ground-truth.
We compare among three methods: (1)
TwitterTrend
, recommending videos related
to the Twitter trending topics; (2)
YouTubeProfile
, recommending videos according to
user YouTube profile; and (3)
T-YRecommender
, recommending based on Eq. (
5.3
).
F-score@
K
is utilized as the evaluation metric for recommending top-
K
YouTube
videos to the target user. Figure
5.4
shows the experimental results. We can see that
Fig. 5.4
Performance comparison
