RBF Neural Network for Feature Selection Using Sparsity Method

The simplicity of RBF neural network (RBFNN) mainly depends on the input (feature) and hidden nodes. To compact the structure of RBFNN, in this paper, a weight decay regularizer based integrated feature selection (FS) strategy is proposed to prune the input nodes. The training procedure of our FS method is explainable: firstly using clustering algorithm, the initialization process of RBFNN is detailedly described; then a novel memory based gradient method is used to promote the process of feature selection and parameter optimization; finally, the weight decay terms for FS will tend to different values. Using two regression problems, the validation of our model to realize feature selection and predict the real outputs is proved.