Efficient All-and-One Support Vector Machines Based on One-versus-All Data Inseparability

摘要

We introduce a new strategy to estimate data inseparabilities between pairs of classes based on One-Versus-All (OVA) classifiers, and use them to enhance the All-And-One (A&O) technique. For an N-class problem, in the worst case the proposed method requires only N+1 binary classifiers to obtain the classification result as same as in the traditional framework. Our proposed method will properly create the necessary set of One-Versus-One (OVO) classifiers corresponding to OVA data inseparabilities in advance. On the other hand, it will not needlessly construct an OVO classifier if the data separability between that pair of classes is good enough. This is more practical than the traditional A&O technique that suggests to train OVO classifiers on-demand which consumes time while the system is in use. Especially in same cases of real-time applications, it may be very important to keep high speed response and maintain highly accurate results. Experimental results show that the proposed technique gives significantly better performance compared to the OVA and the traditional A&O techniques, while providing comparable accuracy compared to the OVO method. Moreover, a required number of binary classifiers in our proposed method is considerably smaller than in the OVO approach as similar as in the OVA and the traditional A&O techniques.