Subgradient methods for the efficient learning of minimax risk classifiers

Machine learning techniques are very important nowadays. The most popular scenario for these problems is the supervised learning one, where a classification task is defined in order to classify new data. The logical approach of defining a classifier is to achieve the one failing less, so that this is equivalent to minimizing an error function. There are some techniques and algorithms used to solve these problems, but the ones used for nondifferentiable functions have a high computational cost. For a specific type of classifiers called Minimax Risk Classifiers, the error function obtained is not differentiable, so we use the theoretical base of different subgradient methods to develop an algorithm for this specific case. The presented techniques allow the optimization problem to be solved efficiently.