An Evolutionary Feature Selection Technique Using Polynomial Neural Network

In this paper we propose a novel approach for feature subset selection by the Polynomial Neural Network (PNN) using Genetic Algorithm (GA). A randomly selected subset of features of a d ataset is passed to the PNN as input. The classification accuracy of PNN is taken as the fitness function of GA. In the conventional PNN approaches, published in literature so far, the processing by PNN takes large computation time due to the expansion of the whole network at different levels. In the proposed scheme, less number of features selected stochastically using the GA, prevents PNN to grow at very early stages which reduces the computation cost as well as time. The proposed scheme is simulated on six benchmark databases and classification accuracies obtained using proposed PNN classifiers are compared with those obtained using three other existing approaches. It is observed that the classification accuracies using proposed scheme are quite satisfactory compared to existing three schemes. The strength of proposed scheme is justified in two ways: (i) its high classification accuracy with much less computational cost in the presence of reduced number of features and (ii) much less execution time taken by it as compared to other schemes.