24 3. MULTIMODAL TRANSDUCTIVE LEARNING
3.4 RELATED WORK
3.4.1 POPULARITY PREDICTION
Significant efforts were devoted to exploring the popularity prediction of items such as text [5,
107], images [16, 54, 112], and videos [18, 85, 100, 156] due to their potential value in busi-
ness [158, 170].
For the task of predicting the popularity of text, most methods tend to explore the textual
content itself and the correlation between popularity and the social context. For example, Ma
et al. [107] proposed to predict the popularity of new hashtags on Twitter by extracting 7 content
features from both hashtags and tweets and 11 contextual features from the social graphs formed
by users.
As to the image popularity, content-based image features, context features, and social con-
text features are generally exploited to predict image popularity. For example, Khosla et al. [77]
explored the relative significance of individual features involving multiple visual features, such
as color, gradient, texture, and the presence of objects, as well as various social context features,
such as the number of normalized views or contacts. Totti et al. [155] presented a complementary
analysis on how the aesthetic properties, such as brightness, contrast and sharpness, and seman-
tics contribute to image popularity. Gelli et al. [54] proposed to combine user features and three
context features together with image sentiment features to better predict the popularity of social
images.
When it comes to video popularity prediction, analogous to images, videos also in-
tegrate different information channels, like visual, acoustic, social, and textual modalities.
e majority of studies focus on investigating the factors that determine the popularity of
videos [18, 85, 156, 173]. For example, Cha et al. [18] conducted a large-scale data-driven analy-
sis to uncover the latent correlations between video popularity and UGC. Li et al. [85] proposed
using both video attractiveness and social context as inputs to predict video views on online
social networks. Trzcinski et al. [156] employed temporal and visual cues to predict the popu-
larity of online videos. e tasks above share the same thing—they do not describe each item
based on its content only; instead, they mine multiple views of context information related to the
item and social cues from the users to improve the prediction performance. In addition, some
researchers [24, 38, 85] have explored the improvement of video popularity prediction by fus-
ing information introduced by different patterns. To overcome the ineffectiveness of traditional
models, such as autoregressive integrated moving average (ARIMA), multiple linear regression
(MLR), and k-nearest neighbors (kNN), when predicting the popularity of online videos, Li
et al. [85] introduced a novel propagation-based popularity prediction method by considering
both video intrinsic attractiveness and the underlying propagation structure. Roy et al. [138]
used transfer learning to model the social prominence of videos, in which an intermediate topic
space is constructed to connect the social and video domains. Ding et al. [38] developed a dual
sentimental Hawkes process (DSHP) for video popularity prediction, which not only takes sen-
..................Content has been hidden....................
You can't read the all page of ebook, please click here login for view all page.