Personalized Multimedia Search - User-centric Social Multimedia Computing - page 61

Fig. 4.2 Probabilistic
generative model of mmTIM.
Reference [ 27 ] c 2013

Association for Computing
Machinery, Inc. Reprinted by
permission

4.3.2 Multimodal Topic-Sensitive Influence Model

Figure 4.2 illustrates our understanding of the generation process as graphical
structure. In graphical models, nodes represent random variables: shaded node is
observation, unshaded node is hidden state, which needs to be inferred from the
observation. Arcs represent independence assumption, which usually correspond to
a sampling operation. In this work, the observations, namely inputs, include the
contact network, tag vectors, and image vectors. To unify image and tag analysis,
we represent image by responses on clustered visual descriptors: we first construct
a visual vocabulary and represent each user by the visual descriptor responses v
of his/her uploaded images. On the right of the model, we intimate the generation
process of image visual content. On the left, we intimate tags' generation.
The generation of image and tag share the same set of latent variables, including
(1) the boolean switch variable, s w and s v , to control whether the tag words and
visual descriptors are generated by the users themselves or according to one of the
influencers; (2) the contact user variable, to record the selected contact user when
s w 0or s v ʩ c w ; (4) topic index variable

z w and z v , to record the selected topic; and (5) the topic-word distribution

= = 1; (3) the user-topic distribution ʩ u and

ʦ , where

w ) or visual descriptor ( v ). When s w

word can be either tag ( ʦ ʦ = 1, the tag word is

ʩ u . When s w

generated based on user's own interest = 0, the tag word is influenced

by one of users in the contact list ʩ c w

of the sampled influencer c w . The details of the generative process for tag words of
user u are as follows:

C u , and it is generated based on the interest

For each tag word w u , i w u ,

(1) Draw a switch variable from a bernoulli distribution: s u , i

∈

∼ Bernoulli (ʻ) ;

(2) If s u , i = 0, then

i. Draw a influencer from u 's contact list, which follows multinominal distri-
bution parametered by ʳ :

c u , i

;
ii. Draw a topic from influencer c u , i 's topic distribution: z u , i

∼ Multi (ʳ )

∼ Multi (ʩ c u , i ) ;