Information Technology Reference
In-Depth Information
5.4.2.3 Results and Analysis
To better understand the association between heterogeneous topics, we examine
the derived association matrix A . Among 80
×
40
=
3,200 association pairs, the
z Yo u
1
z Twi
43
z Yo u
17
z Twi
38
most significant two are
{
,
}
and
{
,
}
, which have been visualized in
Tables 5.2 and 5.3 .
We can see that the derived association involves with multiple aspects: game-
related YouTube topic #1 significantly associates with Twitter topic #43 whose
top-ranked followees are official game platforms or developers, and the associa-
tion between YouTube topic #17 and Twitter topic #38 results from their shared
location in Germany. A single association metric, e.g., semantics, tends to fail in
this case. Actually, one fundamental advantage of exploiting the overlapped users
for association mining is its flexibility: there is no need to explicitly design an asso-
ciation metric. Users' collaborative activities on different social networks implicitly
define the metric.
5.4.3 Cross-Network Application
With the derived topic association matrix A , given any user's YouTube video topic
distribution, we can estimate his/her Twitter followee topic distribution, i.e., the
most probably followed followees on Twitter. In the context of YouTube video pro-
motion, the focus is YouTube video. We view each test YouTube video v as a virtual
YouTube user who holds identical topical distribution v
z Yo u
. It is easy
to understand that the virtual user actually represents the typical users in YouTube
showing significant interest to the test video, who are exactly the potential fans, and
thus the promoted targets. Therefore, after topical distribution transfer, the virtual
user's Twitter followee topic distribution v Twi
=
p
(
|
v
)
z Twi
reflects the most probable
Twitter following patterns for the video fans. It is promising to identify the Twit-
ter followee that best matches this transferred distribution as the optimal promotion
referrer for the video.
At Twitter side, for each popular followee u , his/her Twitter topic distribution
u Twi can be calculated as:
=
p
(
|
v
)
z Twi
k
z Twi
k
z Twi
k
p
(
|
u
)
p
(
u
|
) ·
p
(
)
z Twi
k
where p
(
)
is the topic prior and can be calculated by aggregating over users.
z Twi
k
Here p
(
|
u
)
actually reflects followee u 's popularity in the k th Twitter followee
topic.
Given the test YouTube video and candidate Twitter followees represented on the
same topic space, the direct way is to use dot product as the matching measure. The
matching score of Twitter followee u to promote YouTube video v is calculated as:
 
Search WWH ::




Custom Search