Error prediction for multi-classification

This paper describes an error prediction mechanism for multiclassification systems. First, a multiclassification system is constructed by combining a suite of two-class classifiers. While training, each sub-classifier does not utilize all the training data and the remaining data are used for testing purpose. Thus, the classification system can predict its own performance after training. We have tested this mechanism on several well-known benchmark datasets. Experimental results are demonstrated for its effectiveness.