Classification improvement based on feature combination and topic vector model

In this paper, we demonstrate a feature processing procedure which emphasizes on the combination of original features with redundancy trimming steps. This procedure shows better classification result than traditional classification models. In our experiment, several key feature processing steps were proposed according to the type of the feature. These steps contains numerical to categorical feature value conversion, feature combination, feature redundancy discrimination, and latent structure discovery based on the concatenation of original features and extended feature set. The UCI machine learning repository is chosen as our demonstration to show the effect of our approach. In our preliminary result, it shows that the classification accuracy outperforms the traditional SVM classifier(SVM-only) while the ROC benchmark equals to the SVM-only scenario. This result is believed to be a promising one on the feature processing procedure research.