For multi-view learning, existing methods usually exploit originally provided features for classifier training, which ignore the latent correlation between different views. In this paper, semantic features integrating information from multiple views are extracted for pattern representation. Canonical correlation analysis is used to learn the representation of semantic spaces where semantic features are projections of original features on the basis vectors of the spaces. We investigate the feasibility of semantic features on two learning paradigms: semi-supervised learning and active learning. Experiments on text classification with two state-of-the-art multi-view learning algorithms co-training and co-testing indicate that this use of semantic features can lead to a significant improvement of performance.
Financed by the National Centre for Research and Development under grant No. SP/I/1/77065/10 by the strategic scientific research and experimental development program:
SYNAT - “Interdisciplinary System for Interactive Scientific and Scientific-Technical Information”.