A Joint Evolutionary Method Based on Neural Network for Feature Selection

Feature selection, structure determination and connection weights training are three key tasks for the classification problem based on neural network. Traditional feature selection methods with neural networks neglect the fact that these three tasks are interdependent and make a joint contribution to the performance of neural network, which often results in an irrational network structure and unsatisfying generalization capability. In order to solve the above problem, a joint evolutionary method based on neural network for feature selection is proposed in this paper. A hybrid representation scheme and the crossover operator based on the generated subnet are employed in consideration of the relationship between genotype and phenotype. By introducing penalty factor for the number of input nodes and hidden nodes into fitness function, the input feature subset and the network structure are evolved jointly. The experimental results with three real-world problems show that the proposed method not only accomplishes effectively feature selection but also improves the classification accuracy.