Estudo de funções de custo para redes neurais com dados desbalanceados

Abstract

The work presented here makes a comparative approach of the techniques used in neural networks in problems of unbalanced classes. Based on an initial comparison of classical techniques, a more in-depth study is done under the use of cost functions to deal with the problem during the training phase in neural networks with unbalanced data. An approach about the inclusion of a priori information in the cross-entropy cost function is presented together with a modification of the resilient backpropagation algorithm and the impacts on the learning algorithm. When working with problems of unbalanced classes, measuring the performance of the learning algorithm requires more appropriate metrics such as AUC, F1-score, Kubat's G-mean (Geometric-mean), AGm (Adjusted Geometric-mean) and others. However, the vast majority of problems in this area are trained using the mean square error or cross-entropy (also known as logistic error function). This makes the neural network learning algorithm to optimize a cost function different from the one that will be used to validate its performance. An approach is then presented on how to extract appropriate metrics for this kind of problem from the confusion matrix and transform them into cost functions to be used during the training phase. A comparative study between the traditional training approach and the presented cost functions is carried out, presenting the positives and negatives points of each approach. Numerical experiments for different training bases with different unbalanced rates are presented.