The semiproximal SVM approach for multiple instance learning: a kernel-based computational study

Abstract

The semiproximal Support Vector Machine technique is a recent approach for Multiple Instance Learning (MIL) problems. It exploits the benefits exhibited in the supervised learning by the Support Vector Machine technique, in terms of generalization capability, and by the Proximal Support Vector Machine approach in terms of efficiency. We investigate the possibility of embedding the kernel transformations into the semiproximal framework to further improve the testing accuracy. Numerical results on benchmark MIL data sets show the effectiveness of our proposal.