PF-sGAMCO: A Method for Feature Selection and Neural Network Architecture Design

Neural networks are a commonly used method in machine learning. However, there is no unique critérium to design neural architectures up to now. Feature selection relates to the design of a neural network since it defines the number of inputs which impacts the complexity of its architecture. This paper presents a method that chooses input features by means of a linear filter based on the Pearson correlation coefficient and designs the architecture of a neural network using a structured genetic algorithm with multiple crossover operations. The method was tested over three problems (binary classification, multi-class classification and regression) showing promissory results in comparison with other techniques reported literature.