One of the problems that have to be overcome in classification tasks is high data dimensionality. Therefore, dimensionality reduction techniques such as feature selection have to be employed. Feature selection involves univariate or multivariate evaluation of features with respect to the classification accuracy. Pairwise feature selection was recently proposed as a trade-off between selection process complexity and the need to analyze relationships between features. In our previous work we have proposed a correlation-based modification of the pairwise feature selection.

In this paper we present the results of the experiments in which we have compared the correlation-based feature selection strategy with the unmodified pairwise approach. The experiments were performed using neural network classifiers on commonly used benchmark data sets.