Generating kernel matrix for rotation forest through genetic programming
Mojtaba Khamar, Mahdi Eftekhari · 2018
Classification is one of the most important issues in real world. Recent researches advocate combining multiple classifiers, e.g, ensemble learning methods. These methods are the common approaches for classification that create a set of classifiers and then classify new data points by majority voting. Also, evolutionary algorithms have been used for finding optimal parameters and classifiers in classification issues, e.g, Genetic Programming (GP). In this paper, a new RF method is proposed and called Rotation Kernel Forest (RKF). In RKF method: first, some equations are generated by GP that are employed as the feature function φ(χ). In the second step, kernel matrix is constructed based on φ(χ) and at the end, projection matrix is achieved. RKF method generates not only a new kernel matrix but also a new projection matrix. The experimental results show apparently the efficiency of RKF comparing to the advanced ensemble methods in terms of accuracy of classification. Wilcoxon signed-ranks test confirms the superiority of RKF in comparison to the other methods.