A Transfer-Learning Approach to Exploit Noisy Information for Classification and Its Application on Sentiment Detection - Technologies and Applications of Artificial Intelligence - page 266

3.2 Noise-Label Transfer Learning (NLTL)

We propose NLTL, which is a transfer learning model to solve the above-mentioned
problem. The overall architecture is shown in Fig. 3. The idea is to transfer informa-
tion from low-quality domain data to improve the prediction in high-quality domain
which has insufficient training instances. Note that for each object, we may integrate
corresponding instances from multiple low-quality data sources. NLTL first uses in-
stances existing in both high-quality and low-quality domains as a bridge to identify
the correlation between coarse-grained and fine-grained features. Then it learns the
weight of instances from each domain to train a binary classifier to predict testing
data in the high-quality domain. It should be noted that we perform feature transfer on
both training and testing data, however, only training data are used to learn the weight
of instances since testing data are not labeled. We define NLTL in Algorithm 1. Fea-
ture transfer is performed using Structural Corresponding Learning (SCL) [4] (Step 1
to Step 4, see 3.3), and TrAdaBoost [3] is used to tune the weight of instances (Step 5
to Step 12, see 3.4).

3.3 Feature Transferring

We want to handle the problem that the quality of features in low-quality domain is
not as good as that in high-quality domain in terms of granularity. The goal is to iden-
tify a mapping function to project the features in the low quality domain to the high
quality domain, by changing their distributions.
We propose a method based on Structural Corresponding Learning (SCL) [4]. The
intuition is to identify the correspondences among the features from different domains
by modeling their correlation with features that have similar distribution in both do-
mains. To transfer the low-quality data into high-quality domain, for each feature in
the low-quality domain, it is necessary to find its mapping to the more fine-grained
high-quality domain. Here we propose to create a prediction model to perform the
mapping. That is, for each feature in the high-quality domain, we create a classifica-
tion or regression model, for categorical and numerical features respectively, to pre-
dict its value given each corresponding instance in the low-quality domain. Assume
an user u appears in both high-quality domain (its feature vector, denoted as , is
{“Male”, “22”, “May”, “Taipei”, “Software Engineer”} ) and low-quality domain
(feature vector denoted as , which is {“Male”, “20 to 30”, “May”, “Taiwan”,
“Engineer”}). will of course be used as the training example to learn a compul-
sive user model, but we want to use as well to enlarge the training set. Therefore,
for each feature in the high-quality domain, we create a classifier that maps to a
corresponding value. In our example, we will build 4 classifiers and 1 regressor (for
'age' feature), each of which takes an instance in as input and output the possible
assignment for the fine-grained feature.
We denote these models as mapping function , and it models the correlation be-
tween the features from different domain. In the experiment we use linear regression
to learn .

· · ·