Decision confidence-based multi-level support vector machines

Support vector machines (SVM) have been showing high accuracy of prediction in many applications. However, as any statistical learning algorithm, SVM's accuracy drops if some of the training points are contaminated by an unknown source of noise. The choice of clean training points is critical to avoid the overfitting problem which occurs generally when the model is excessively complex, which is reflected by a high accuracy over the training set and a low accuracy over the testing set (unseen points). In this paper we present a new multi-level SVM architecture that splits the training set into points that are labeled as 'easily classifiable' which do not cause an increase in the model complexity and 'non-easily classifiable' which are responsible for increasing the complexity. This method is used to create an SVM architecture that yields on average a higher accuracy than a traditional soft margin SVM trained with the same training set. The architecture is tested on the well known US postal handwritten digit recognition problem, the Wisconsin breast cancer dataset and on the agitation detection dataset. The results show an increase in the overall accuracy for the three datasets. Throughout this paper the word confidence is used to denote the confidence over the decision as commonly used in the literature.