A Transfer-Learning Approach to Exploit Noisy Information for Classification and Its Application on Sentiment Detection - Technologies and Applications of Artificial Intelligence - page 269

Fig. 4. Framework of Sentiment Diffusion Prediction with NLTL

The results show that NLTL outperforms the competitors for all dataset, especially
for CTG. It also shows that by exploiting low-quality domain data, NLTL is useful
and can improve the result using only high-quality domain data (denoted as High-
Quality in Table 1) up to 6.7% in terms of AUC. On the other hand, NLTL combines
the advantages of TrAdaBoost and SCL. It considers not only features but also labels
to the same items together. The improvement of NLTL over baseline algorithms
shows that both features and labels information from low-quality domain data are
important and useful.

5 Sentiment Diffusion Prediction on Novel Topics

In this section, we use NLTL to handle a novel real-world sentiment diffusion predic-
tion problem. Sentiment prediction aims at predicting whether an opinion is positive
or negative [6]. However, in this application, we are interested in predicting the diffu-
sion of sentiment through social networks. In other words, we emphasize on senti-
ment “diffused” rather than sentiment “expressed” by a user. Analyzing sentiment
diffusion allows us to understand how people react to other people's comments on
micro-blog platforms.
Traditional sentiment prediction uses a variety of textual or linguist information as
features [6]. Such solution has a serious drawback as it is unable to handle new topics
that appear rarely. On the other hand, Kuo et al. [7] propose a method to predict the
diffusion on novel topics utilizing latent and social features. Rather than predicting
the existence of diffusion, we extend [7] to predict the diffusion of sentiments.