Divide & Conquer Classification and Optimization by Genetic Algorithm

In this paper, a new approach for improving the performance of recognition system is proposed. The main idea of proposed approach is using pairwise classifiers. Firstly, a multi class classifier is trained and its confusion matrix is derived. Then, the error between metaclasses is derived. In each level, our objective is to minimize the error between metaclasses in the evaluation dataset. This method is similar to creation of a binary tree. Each time the data is divided into two metaclasses, until there is no node greater than one class. Each node is equal to one classifier that distinguishes the classes of the left and right nodes. The genetic algorithm makes sure that we have the minimum error in confusion matrix. The Multi Layer Perceptron and K-Nearest Neighbor are used as base classifiers. Experimental results demonstrate improved accuracy on a Farsi digit handwritten dataset.