Research On Support Vector Regression In Prediction Of Infectious Diseases

The threat to human from infectious diseases has been tough for a long time. Forecasting can detect the infectious diseases epidemic trends, establish foundation for early warning and be significant for infectious diseases control.Affected by many factors, the epidemic incidence samples are difficult to collect,and always present irregular, chaotic and other nonlinear characteristics. However, most of the traditional prediction models are linear, which are inappropriate for nonlinear fitting. Thus, it's an urgent need to study new models for infectious diseases forecasting. Taking the advantages of the support vector machine (SVM) into account, such as small samples, sparse, good nonlinear fitting ability and so on, this thesis introduces the support vector regression (SVR) into the epidemic prediction, which aims to establish superior forecasting models and provide new technological means for diseases prediction.Firstly, this thesis summarizes the current study of infectious prediction and SVM, and outlines some representative epidemic mathematical models. With an overview on the statistical learning theory and the related optimization theory, we expound the basic knowledge of SVR model and point out the advantages and disadvantages of SVM.Secondly, this thesis designs the NSVR prediction model in allusion to the difficulties in practical application of SVM, which aims to achieve a higher performance and prediction accuracy. The designed ideas relate to three main facets: (1) data preprocessing: phase space reconstruction by C-C method; (2) the choice of kernel function: adopt the nonlinear combination of mixed kernel function; (3) parameters optimization: improve the basic PSO algorithm. On the basis of describing the modeling process of NSVR prediction model, the result on the emulational experiment shows that the model is effective.Finally, this thesis proposes a new combination forecasting model ARIMA-NSVR with the idea of combined prediction, in related to the advantages of the autoregressive integrated moving average (ARIMA) model in dealing with linear problems and SVR in dealing with nonlinear problems, which aims to further improve the robustness and the generalization ability. On the basis of the modeling process of this combined prediction model, the result on the emulational experiment shows that the model is effective.