A Transfer-Learning Approach to Exploit Noisy Information for Classification and Its Application on Sentiment Detection - Technologies and Applications of Artificial Intelligence - page 271

Table 2. Sentiment Diffusion Prediction results

All Features Best Features

High-Quality 62.90% 65.04%

Low-Quality_1 61.73% 65.36%

Low-Quality_2 63.47% 66.26%

All Instance 62.13% 64.25%

61.65% 62.33%

SCL

61.84% 65.27%

TrAdaBoost

59.58% 62.59%

Label-Powerset

NLTL 64.21% 68.30%

• User Sentiment Information. Similar to link sentiment information, user senti-

ment information models the tendency of each user to reply to positive/negative
posts. For a user, we generate the user sentiment score according to sender aspect
( ), receiver aspect ( ), and sender-receiver aspect ( ). More specifical-
ly, for we only consider the number of positive and negative posts sent by us-
er, and ignore those received by this user. On the other hand, only considers
the number of positive and negative posts received by user. considers both
aspects.

• Topic Information. We follow the same approach described in [7] to extract latent

topic signature ( ) features. Besides TG , we also extract topic similarity ( TS )
features weighted by link sentiment information and user sentiment information.
There are four features generated based on topic similarity, topic similarity for link
sentiment ( ), topic similarity for user sentiment with sender aspect ( ),
topic similarity for user sentiment with receiver aspect ( ), and topic similar-
ity for user sentiment with sender-receiver aspects ( ).

• Global Information. We extract global social features such as in-degree ( ID ), out-

degree ( OD ), and total-degree ( TD ) from social network. Note that these three fea-
tures remain the same for different labeling methods; thus, we utilize them as pivot
features in SCL and NLTL algorithms.

5.4 Results

The experiment setting of sentiment diffusion prediction task is the same as that de-
scribed in Section 4. We compare NLTL that utilizes three sources to the competi-
tors as described in 4.1. We run the experiment on two set of feature combinations:
using all features and the best feature combination chosen using wrapper-based for-
ward selection method [9]. The result shows that NLTL is able to integrate the infor-
mation of features and labels to outperform the competitors by a large margin.