Multi-Class Classification based on Relative Distribution of Class

Binarization of multi-class classification problem into two class problem is widely adopted in machine learning because of its simplicity and efficiency. It consists of dividing multiple classes into pairs of all possible combinations and learning the base classifiers on each pair of classes. Then, their outputs are combined to classify an instance. To improve the classification accuracy, several different combination schemes were studied previously. We proposed a new combination scheme based on relative distribution of each class. Instead of merely computing the distances of an instance to the nearest neighbors of each class, relative distances were measured considering the relative distribution of each class. Experimental results showed the proposed method outperforms previous methods both in terms of accuracy and kappa measures.