Information Technology Reference
In-Depth Information
actually indicates users' temporal response patterns between the two networks. With
“T-W ratio” larger than one, we can see that in all the examined topics, majority
of overlapped users first involve into the trending topics on Twitter. This concludes
that the observation of “Twitter is faster than YouTube” also applies at the micro
user level.
5.3.2 YouTube Video Recommendation
The observation derived from the above data analysis opens up possibility to exploit
the heterogeneous social multimedia data associated to the same overlapped user.
Since Twitter response is ahead of YouTube at user level, for specific user, we
can employ the observed Twitter activity as auxiliary knowledge to predict his/her
YouTube activity and guide personalized YouTube video recommendation.
Recently, short-term interest modeling has attracted attentions in personalized ser-
vices [
28
]. Compared with long-term interest, short-term interest changes dynam-
ically and is more vulnerable to the trending topics. Therefore, a key problem in
short-term interest modeling is how to obtain real-time and abundant user activity
data. Our solution is to exploit the Twitter activity to make up for the data shortage
in YouTube. The basic premise is that, if user is discovered to get involved into a
trending topic on Twitter, it is very likely he/she will have subsequent activity on
YouTube in a short time and we are confident to recommend him/her the related
YouTube videos.
Based on that, we develop a personalized YouTube video recommender by inte-
grating cross-network user social activities. For specific overlapped user, the Twitter
activity and YouTube activity data are utilized to extract the short-term and long-term
interest, respectively. The final personalized recommendation list is then obtained by
considering both short-term and long-term interests (as illustrated in Fig.
5.3
). Since
Twitter is recognized as discovering the trending topics [
23
], in this way, we can
quickly capture users' interest in the emerging hot topics, and recommend timely
and interesting videos to them on YouTube. Since most related videos to the trending
topics are newly uploaded, traditional popularity-based recommendation strategies
will fail in discovering these videos or recommending them to the desired users.
5.3.2.1 Short-Term User Modeling
We have employed straightforward techniques to realize the solution. For short-term
user modeling from Twitter, we modify Twitter-LDA [
33
], which is an extension
to the standard Latent Dirichlet Allocation (LDA) by addressing the 140-character
challenge. In Twitter-LDA, each document (tweet or retweet) is assumed to be gen-
erated from one single topic and one background topic, where the background topic
consists of frequent words shared between documents. In our problem, users' short-
term interest usually highly correlates with the trending topics. Within a short time
