Design of prediction model for multiple feature-type problems based on improved PSO-RBF neural network

Prediction problems for complex systems have been of wide interest. Since the input features of such problems often contain both qualitative and quantitative variables, it is not possible to preprocess the features by conventional methods. Moreover, due to the complexity of the problem, using only a single model for prediction may lead to insufficient accuracy. In this paper, a prediction model for multiple feature-type problems based on an improved PSO-RBF neural network is designed, in which qualitative and quantitative features are processed separately. The qualitative features are first screened using information gain and then the quantitative features are downscaled using factor analysis. After processing the features, they are predicted using improved PSO-RBF neural network. The results show that the resulting optimal SSE is 77136 and the prediction accuracy is high.