A Multi-view Learning Method Based on Feature Correlation

The multi-view learning can improve the generalization performance of the classifier by using features from different sources. However, many multi-view data sets are constructed by unsupervised methods, which may cause uncertain to the classifier. In this paper, a supervised multi-view learning method based on the correlation between features is proposed, which mainly includes two parts: view evaluation and view construction. First of all, canonical correlation analysis and correlation matrix based on mutual information are employed to extract the correlation information between different feature sets. After that, a supervised data view construction algorithm, named data view construction (DVC), is proposed. Compared with traditional view construction methods, DVC can be used as a general view construction method to construct multi-view data set. The experimental results show that DVC achieves higher generalization compared with other view construction algorithms.