Enhancing Multi-Class Classification in One-Versus-One Strategy: A Type of Base Classifier Modification and Weighted Voting Mechanism

An effective way to solve multi-class classification problems is the One-Versus-One (OVO) decomposition strategy. When applying the OVO strategy to practice, there may be a `tie vote' phenomenon so that some samples cannot be classified. This article proposed a weighted voting mechanism formulated by base classifier accuracy, which is simple and effective to eliminate the unclassifiable area. In addition, one can further improve the classification ability of base classifiers by modifying two-class base classifiers to three-class base classifiers. In the simulation experiments, logistic regression and support vector machine are taken as base classifiers, and the white wine dataset and red wine dataset are viewed as the benchmark of four-class and five-class classification respectively. The results indicate that for multi-class classification, higher accuracy and efficiency can be achieved by modifying the basic classifier and implementing the weighted voting mechanism.