Information Fusion based on Multiplicity of Data Preprocessing Boosts the AdaBoost Classifier

The paper considers the possibility of boosting efficiency of the AdaBoost algorithm by introducing diversity in the information extraction procedures and subsequent information fusion. A given raw data set is transformed by two well known methods known as vector autoscaling and dimensional autoscaling. These methods transform data by mean-centering and variancenormalization with respect to the measurement variables and the measured samples, respectively. The feature (information) extraction is implemented by the principal component analysis. Final data representation is obtained by two alternate fusion strategies. In one, feature extraction is done immediately after the scaling transformations and the two feature sets are fused by a simple procedure of concatenation. In the other, data vectors obtained by the two scaling transformations are first fused by simple concatenation and then the feature extraction is done. The data representation is thus by fused feature vectors of dimensionality twice that of the original data. The AdaBoost algorithm by Freund and Schapire is implemented by using a simple threshold base classifier. The classification efficiency of these procedures is compared with that obtained by using single preprocessor based AdaBoosting. Four benchmark data sets are used for validation. The analysis demonstrates that the performance of the AdaBoost algorithm is enhanced further by multiple 170 Divya Somvanshi and R.D.S. Yadava preprocessor based fusion of data space, particularly if the original variables are of same kind.