Growing Support Vector Classifiers with controlled complexity
Semiparametric Support Vector Machines have shown to present advantages with respect to nonparametric approaches, in the sense that generalization capability is further improved and the size of the machines is always under control. We propose here an incremental procedure for Growing Support Vector Classifiers, which serves to avoid an a priori architecture estimation or the application of a pruning mechanism after SVM training. The proposed growing approach also opens up new possibilities for dealing with multi-kernel machines, automatic selection of hyperparameters, and fast classification methods. The performance of the proposed algorithm and its extensions is evaluated using several benchmark problems.