Personalized Celebrity Video Search Based on Cross-space Mining


Zhengyu Deng, Jitao Sang and Changsheng Xu*



(1)  The videos about celebrities are closely followed by users because of the “Celebrity Effect”.
(2)  Celebrities are often popular in multiple fields and users’ interests are diverse. To a specific user, he/she often only take interest in certain field of a celebrity.
(3)  We present an method to match user interest distribution and celebrity popularity distribution to realize personalized celebrity video search.






For the celebrity side, celebrity popularity is explored by leveraging expert information, e.g., the corresponding wikipedia homepages. Standard topic modeling method of Latent Dirichlet Allocation (LDA) is adopted to extract the celebrity popularity distribution in abstract topic level. For the user side, since off-the-shelf user profile is unavailable or hardly informative, we exploit user interest based on his/her online activities, e.g., video sharing, social tagging. LDA is again utilized for user interest topic extraction. Given the derived heterogeneous popularity and interest spaces, we introduce a cross-space correlation method. Semantic and context intra-word relations are refined by random walk to bridge the interest and popularity spaces.


Cross-space Correlation



  The illustration of cross-space correlation is shown above.




Experimeantal settings

In order to evaluate the performance of our approach, we compare with 1) the method that just learn a united topic space for users and celebrities and 2) non-personalized search. The performance assessment measure is F-score.



Experimental Results

Parts of the discovered latent topics of interest and popularity spaces are displayed in the following table, which confirms our hypothesis that the latent topics of user and celebrity could be well extracted via LDA.



The comparison of average F-score at different depths is illustrated in the following figure (the number of latent topics and the weight of random walk is tuned to its optimal value.)



The influence of random walk is shown in the following figure.






Personalized Celebrity Video Search Based on Cross-space Mining [pdf] [slides] [poster] [code] [data]

Zhengyu Deng, Jitao Sang and Changsheng Xu
In Pacific Rim Conference on Multimedia (PCM), Singapore, Dec. 2012, pp.455-463.