Feature selection using Double Parallel Feedforward Neural Networks and Particle Swarm Optimization

In recent years, the neural network (NN) based feature selection becomes a promising method for dimensionality reduction. However, multi-layer feedforward neural network (MFNN) with wide applications has some disadvantages such as local minimal points on the error surface and over-fitting problem. At the same time, the conventional approaches usually fixing the number of hidden nodes and focusing on the input selection hinder further remove of the redundant information and improvement of network generalization performance. To solve these problems, a feature selection algorithm using double parallel feedforward neural network (DPFNN) and particle swarm optimization (PSO) is proposed. The algorithm adopts DPFNN with the merits of single-layer feedforward neural network (SFNN) and MFNN as the criterion function, synchronously performs optimization of structure and selection of inputs based on a new defined fitness function keeping balance between network performance and complexity. Experimental results show that the algorithm can effectively remove the redundant features while improving the generalization ability of network.