Subspace Selective Ensemble Algorithm Based on Feature Clustering

A feature-clustering-based subspace selective ensemble learning algorithm was proposed to improve ensemble classifier performance, allowing for high dimensional data sets. First, features were clustered on weighted average linkage method and reduced subspaces were generated by extracting an attribute from each feature cluster. Then the feature reduced subsets served as inputs of individual GA-SVMs which had high accuracy to ensure individuals with significant diversities. Some individuals with both diverse and accurate were selected to construct ensemble system. Finally, In Matlab 2010a environment, the algorithm was simulated on 4 datasets. The kappa-error diagrams demonstrated that individual classifiers were both accurate and diverse, and the results showed the classification accuracy increase significantly.