Feature Reduction Using Ensemble Approach

The performance of many content analysis methods heavily dependent on the features they are applied. A fundamental problem that makes the content analysis difficult is the curse of dimensionality. In this study, we propose a novel feature reduction method which adopts ensemble approach to measure the divergence between the training set and test set and use the divergence to supervise the feature reduction procedure. The proposed method uses pairwise measure to get the diversity between classifiers and selects the complementary classifiers to get the pseudo labels on test set. The pseudo labels are used to measure the divergence between training set and test set. The feature reduction algorithm merges the adjacent feature space according to the divergence, such reduce the feature number. We evaluated the proposed method on several standard datasets. Experiment results shown the efficiency of the proposed feature reduction method.