Towards minimizing the energy of slack variables for binary classification

This paper presents a binary classification algorithm that is based on the minimization of the energy of slack variables, called the Mean Squared Slack (MSS). A novel kernel extension is proposed which includes the withholding of just a subset of input patterns that are misclassified during training. The later leads to a time and memory efficient system that converges in a few iterations. Two datasets are exploited for performance evaluation, namely the adult and the vertebral column dataset. Experimental results demonstrate the effectiveness of the proposed algorithm with respect to computation time and scalability. Accuracy is also high. In specific, it equals 84.951% for the adult dataset and 91.935%, for the vertebral column dataset, outperforming state-of-the-art methods.