Improving the Accuracy of Ensemble Classifier Prediction Model Based on FLAME Clustering with Random Forest Algorithm

Abstract

Recent approaches in the area of ensemble classification of data aim to make base classifier's error uncorrelated as possible though learning is given little importance. The substantial increase in the learning of the base classifier can propagate better prediction to the final fusion classifier. Therefore a novel approach to enhance the learning capability of the base classifier by fuzzy based clustering has been proposed in this paper. The learning of the base classifier has been drastically improved with the advent of fuzzy decision boundaries manipulated by the algorithm FLAME known as fuzzy clustering by local approximation of membership of the data in the clusters. The proposed model is a combination of unsupervised and supervised learning. Decision trees are used as the base classifiers which are integrated over the probability model based on Bayes' theorem. Decision trees form the ensemble and fusion classification is performed by the Random Forest algorithm along with Bayesian model averaging. The accuracy is evaluated over benchmark dataset from the UCI machine repository.