Incorporating prior information into Support Vector Machines in the form of ellipsoidal knowledge sets

This paper investigates a learning model in which the training set contains prior information in the form of ellipsoidal knowledge sets. We handle this problem in a minimax setting, which consists of maximizing the worst-case - minimum - margin between the knowledge sets from the two classes and the decision surface. The problem is solved using an alternating optimization scheme and an active learning strategy, i.e., the training set is created progressively according to the prior information. Our approach is evaluated on toy examples and on a usual benchmark database. It is successfully compared to state-of-the-art techniques.