Semi-supervised training of Least Squares Support Vector Machine using a multiobjective evolutionary algorithm

Support Vector Machines (SVMs) are considered state-of-the-art learning machines techniques for classification problems. This paper studies the training of SVMs in the special case of problems in which the raw data to be used for training purposes is composed of both labeled and unlabeled data - the semi-supervised learning problem. This paper proposes the definition of an intermediate problem of attributing labels to the unlabeled data as a multiobjective optimization problem, with the conflicting objectives of minimizing the classification error over the training data set and maximizing the regularity of the resulting classifier. This intermediate problem is solved using an evolutionary multiobjective algorithm, the SPEA2. Simulation results are presented in order to illustrate the suitability of the proposed technique.