Multi-class learning with specific features for pairwise classes

Support vector machine is initially developed for binary classification problem. Multiclass support vector machine (MSVM) is usually realized by using a combination of several binary SVMs. In most of the existing MSVM approaches, all binary SVMs operates on the same feature space. This paper proposed a new approach in which each binary SVM is associated with a specific feature representation. Based on the idea, we developed an algorithm for MSVM named REAL. In the experiment its performance is compared with traditional approaches on 17 real-world multi-class datasets. The good performance achieved by the algorithm clearly verifies the effectiveness of this approach.