Abstract—This paper presents a classifier model based on Rotation Forest (RF) ensemble structure for biomedical data classification. Classifiers based on RF are generally implemented by using Decision Trees. In this study, optimized Neural Network (NN) is preferred as being the base classifier in RF so as to achieve higher classification performance. Two optimization techniques, Artificial Bee Colony Optimization (ABC) and Particle Swarm Optimization (PSO), are utilized to improve the performance of NN for escaping from local minima. In this way, PSO-NN and ABC-NN based RF structures are designed, and they are called as RF (PSO-NN) and RF (ABC-NN), respectively. In these classifiers, initial weights of NNs are found by using PSO or ABC algorithms. The implemented classifiers based on RF are applied to biomedical datasets (Wisconsin Breast Cancer and Pima Indian Diabetes) that are taken from UCI Machine Learning Repository. Furthermore, fourteen different ensemble structures are generated using these algorithms to prove the superiority of the proposed method. When the results are examined using several performance metrics, it is seen that RF (ABC-NN) classifier achieves to more reliable and better results than other classifiers.
Index Terms—Rotation forest, particle swarm optimization, artificial bee colony optimization, neural networks, biomedical data classification.
The authors are with the Electrical & Electronics Engineering Department, Selçuk University, Konya, Turkey (e-mail: hasankoyuncu@selcuk.edu.tr, rpektatli@selcuk.edu.tr).
Cite: Hasan Koyuncu and Rahime Ceylan, "RF Ensemble Novelties Based on Optimized & Backpropagated NNs," International Journal of Machine Learning and Computing vol. 7, no. 4, pp. 76-84, 2017.