Automatic view composition for improving co-training

In this paper, we propose a view composition method for co-training. In order to compose views properly, two assumptions should be satisfied. One is that two views are class-conditionally independent on each other; the other is that the classification information between labels and view is high. We apply Class-Conditional Independent Component Analysis (CC-ICA) to obtain new features which are mutually independent, and compose views hold a high classification information. We show that our method is promising and effective through the experiment.