Multiclass classification with pairwise coupled neural networks or support vector machines

Support Vector Machines (SVMs) are traditionally used for multi-class classification by introducing for each class one SVM trained to distinguish the associated class from all the others. In a recent experiment, we attempted to solve a K-class problem using a similar decomposition with K feedforward binary neural networks. The disappointing results were explained by the fact that neural networks suffer from datasets with a strongly unbalanced class distribution. By opposition to one-per-class, pairwise coupling introduces one binary classifier for each pair of classes and does not degrade the original class distribution. A few papers report evidences that pairwise coupling gives better results for SVMs than one-per-class. This issue is revisited in this paper where one-per-class class and pairwise coupling decomposition schemes used with both, SVMs and neural networks, are compared on a real life problem. Various methods for aggregating the results of pairwise classifiers are experimented. Beside our online handwriting application, experiments on some databases of the Irvine repository are also reported.