Uma abordagem evolucionária para aprendizado semi-supervisionado em máquinas de vetores de suporte

Abstract

The Semi-Supervised Learning paradigm is highly adequate for a class of problems with growing relevance in the context of Machine Learning: those in which there is a large unbalance between the training and the test data sets due to, among other things, the high cost of a classifier. In such class of problems, one cannot ensure that the labeled patterns appropriately represent the system to be learned, limiting the applicability of the Supervised Inductive paradigm. The unlabeled patterns are then used as an additional source of information about the problem being solved, providing increased generalization ability to the achieved solution. The Support Vector Machines (SVMs) are Artificial Neural Networks widely accepted among the Computational Intelligence community. The formulation based on the Statistical Learning Theory and on the separating margin maximization provides the SVMs with extremely high generalization ability. The TSVMs (Transductive Support Vector Machines) extend the SVMs formulation to the context of Semi-Supervised Learning. However, the search for the set of labels that maximize the separating margin between both the training and the test data is therein performed through an exhaustive local search. The non-optimality of such process motivates the development of the GA3SVMs (Genetic Algorithm Semi-Supervised Support Vector Machines), proposed in this piece. An Evolutionary Algorithm is introduced in the search for the optima classifications for the test patterns, inducing a solution with maximum separating margin and high generalization ability. A modified mutation operator, based on the k-Nearest Neighbors transductive method, is also presented, which adds information to the search process and speeds up convergence significantly for the used Genetic Algorithm. Obtained results show the superiority of the proposed approach compared to the traditional TSVMs, for the class of problems studied.