Topic-sensitive Influence Modeling and Its Applications in Social Multimedia Services
Jitao Sang and
Changsheng Xu*
Summary
(1) Large-scale social media data has provided opportunities to multimedia research and applications.
(2) Social relation analysis is important for social media applications in multimedia.
(3) Social relation includes group-wise & peer-to-peer. Peer-to-peer further divides into two-way & one-way.
(4) In social media, typical two-way relation is "friendship", like "connect" in LinkedIn; typical one-way relation is "influence", such as "follow" in Twitter, "contact" in Flickr.
(5) We focus influence analysis of Flickr in this work, and exploit it for application of personalized search.
Framework
For multimedia application, influence affects someone on behaviors and decisions. We claim that fixed binary or continuous influence is not enough, and influence needs to be topic-sensitive.
The framework is divided into two stages: influence modeling and application. For influence modeling, we simultaneously obtain: (1) topic space (2) user exper- -tise distribution and (3) topic-sensitive influence. For application, we employ derived influence to social network-base personalized image search. Task (query) adaptive is realized.
Topic-sensitive Influence Modeling
Assumption
Inspired by the above results of data analysis in tracking interest change after adding contact user, we assume that user tagging and uploading in two ways:
(1) Innovative:create content based on own interest;
(2) Influenced: data generation is affected by contact users.
Generation process
(1) Draw switch variable from Bernoulli distribution:s^w ~ Bernoulli(\lambda)
(2) If s^w=0, (a) Draw influencer from u's contact list: c^w ~ Multinomial(\gamma); (b) Draw topic from c^w's topic distribution: z^w ~ Multinominal(\Psi^w).
(3) If s^w=1, Draw topic directly from u's own topic distribution: z^w ~ Multinominal(\Psi^w).
(4) Draw word from topic-word distribution:w ~ Multinominal(\Phi^w).
Model Learning (Please refer parameter estimation to the paper.)
Discovered topic illustration
Qualitative case study
(1) The identified influencers have high #follower and show strong expertise on the corresponding topics.
(2) The proposed mmTIM shows its capability in identifying the most topic-sensitive influential contact users.
Quantitative evaluation
We compare with two topic-level influence analysis methods designed for text-based networks. Shown in the figure below, mmTIM consistently outperforms the two baselines..
Applications
Risk minimization-based personalized image search
Extend risk minimization framework to personalized image search:
(1) Define Language Models on the derived topic space;
(2) Consider user into query LM and risk calculation;
Experimental Results
Evaluation results on applications of personalized image search and topic-based image recommendation are shown in the figure below. (a) demonstrates the advantage of topic-sensitive over fixed and no influence. (b) validates our motivation that more accurate influence modeling contributes to better application performance.
Publication
Right Buddy Makes the Difference: an Early Exploration of Social Relation Analysis in Multimedia Applications [pdf]
[presentation]
[poster]
Jitao Sang, Changsheng Xu
In ACM Multimedia (MM), Nara, Japan,,
Oct. 2012, pp.19-28.
Social Influence Analysis and Application on Multimedia Sharing Websites [pdf]
Jitao Sang, Changsheng Xu
In ACM Transactions on Multimedia Computing, Communications and Applications (TOMM), Vol. 9, No. 1s, Article 53, 2013.