Implementing reliable learning through Reliable Support Vector Machines

Abstract

Starting from the theoretical framework of reliable learning, a new classification algorithm capable of using prior information on the reliability of a training set has been developed. It consists in a straightforward modification of the standard technique adopted in the conventional Support Vector Machine (SVM) approach: the knowledge about reliability, encoded by adding a binary label to each example of the training set (asserting if the classification is reliable or not), is employed to properly modify the constrained optimization problem for the generalized optimal hyperplane. Hence, the name Reliable Support Vector Machines (RSVM) is adopted for models built according to the proposed algorithm. Specific tests have been carried out to verify how RSVM performs in comparison with standard SVM, showing a significant improvement in classification accuracy.