Performance Analysis of Classifier Fusion Model with Minimum Feature Subset and Rotation of the Dataset

In this paper we have investigated three aspects of classifier fusion system applied to the gender classification problem. To find the minimum subset of features for which the classifier fusion system has better performance, comparison of the three classifier fusion models and the effect of training to testing ratio on the overall output of the classifier fusion system. The final classification results are derived from three models. 1. single best model, 2. fixed combiner, 3. classifiers as combiner model are compared. We have represented the data in the DCT (discrete cosine transform) domain and features extracted through backward search are given to the classifiers. Evaluation of the classifiers and the fusion models are performed on the basis of receiver operating curve (ROC) and AUCH (area under the convex hull). Our findings are that in the fixed combiner fusion model the majority voting rule has the best performance. In single best model KNN (k-nearest neighbor) is the best classifier. The classifier combiner model has best result for using the minimum set of features.