A training set reduction algorithm for feed-forward neural network using minimum boundary vector distance selection

This paper presents an alternative algorithm that can reduce the size of a training set for a feed-forward neural network while obtaining similar levels of accuracy of classification as when the whole training set is used. The algorithm processes the training set by selecting only input vectors at the boundary between consecutive classes of data to be included in the reduced training set. The algorithm finds an input vector of any other class that has minimum distance to each original input vector. This input vector, called paired vector, is one of the representatives of the boundary of its class that constitutes the shape or distribution model of the problem. The algorithm is applied to all input vectors. Only the unique paired vectors are included the reduced training set. No feature or dimension of the training set is sacrificed in the reduction process. The experiments are conducted on a four-class synthetic problem and the six real-world datasets from the UCI Machine Learning Repository. The experimental results present that the algorithm can significantly reduce the size of a training set while obtaining similar levels of accuracy of classification as when the whole training set is used.