Analysis of Methods and Strategies for Diversity Based Genetation of Classifier Ensemble

A classifier ensemble is a group of individual base classifiers. Each classifier is trained individually by modifying the given data set to achieve diversity. During the testing phase the results given by each classifier are collected to give the final result using a technique called as majority voting. Empirical results prove that diversity amongst the base classifiers improves the accuracy of the result. The diversity can be achieved either based on datasets used for training or based on attribute selection methods. In this paper we summarize the study of the ensemble generation methods. We also propose to test a different approach that will find the distance based diversity of randomly generated datasets before the ensemble generation which will not only optimize the construction process but will improve the efficiency also. The Random Forest (RF) is a widely used algorithm for generating ensemble of decision tree classifiers. During the training phase it uses bagging technique to create bootstrapped samples. The best way to test the model is using 10-fold cross validation. As the empirical results show that the diversity in the classifiers improve the overall accuracy of the ensemble, it is needed to verify how diverse the training samples are. This will verify the existence of the base classifier in the ensemble based on the nature of randomly sampled dataset. The base classifier we yield with this approach certainly maintains a higher diversity and the accuracy of the ensemble improves further with these diverse base classifiers. We aim to achieve higher diversity with less number of base classifiers retained in the model.