文摘
One of the most important problems in probabilistic neural network (PNN) operation is the minimization of its structure. In this paper, two heuristic approaches of PNN’s pattern layer reduction are applied. The first method is based on a k-means clustering procedure. In the second approach, the candidates for the network’s pattern neurons are selected on the basis of a support vector machines algorithm. Modified models are compared in the classification problems with the traditional PNN, four well-known computational intelligence algorithms (single decision tree, multilayer perceptron, support vector machines, k-means algorithm) and PNN trained by the state-of-the-art procedures. Seven medical benchmark databases are investigated and one authors’ own real ovarian cancer data set. Comparison is performed on the basis of the global performance indices which depend on the accuracy, sensitivity and specificity. These indices are computed using the standard tenfold cross-validation procedure. On the basis of the reported results, we show that the algorithm based on k-means clustering is a better PNN structure reduction procedure. Furthermore, this algorithm is much less time-consuming.
