An Approach for learning from small and unbalanced data sets using Gaussian noise during artificial neural network training

摘要

Classifiers have been widely applied to scientific and industrial domains, in general, attaining good performance. However, when applied to problems with small and unbalanced data sets, most classifiers have much worse performance, mainly due to the lack of learning ability to cope with these data constraints. The small size of the data set reduces the generalization power of the classifier and the unbalance in class representation makes the classifiers favor the larger and less important classes. Inherent to several real-world problems, small and unbalanced data sets are the reality to be tackled by learning algorithms for producing accurate and reliable classifiers to solve these problems. This paper proposes an approach to address these data constraints based on the addition of Gaussian noise to the input patterns' variables during the training process. The paper focuses on this noise addition to the multi layer perceptron (MLP) neural network (NN). Experimental results indicate that the proposed approach achieves a performance statistically better than the traditional methods (95% confidence) together with lower variability.