Graph theoretic techniques for pruning data and their applications

In pattern recognition tasks, we usually do not pay much attention to the arbitrarily chosen training set of a pattern classifier beforehand. This correspondence proposes several methods for pruning data sets based upon graph theory in order to alleviate redundancy in the original data set while retaining the original data structure as far as possible. The proposed methods are applied to the training sets for pattern recognition by a multilayered perceptron neural network (MLP-NN) and the locations of the centroids of a radial basis function neural network (RBF-NN). The advantage of the proposed graph theoretic methods is that they do not require any calculation for the statistical distributions of the clusters. The experimental results in comparison both with the k-means clustering and with the learning vector quantization (LVQ) methods show that the proposed methods give encouraging performance in terms of computation for data classification tasks.