Research On Ensemble Learning Algorithm Of Incremental NIR Semi-Supervised SVR

The Near infrared (NIR) signal,which represents the composition and concentration of the material, is widely used in agriculture, food, pharmaceutical, medical, petrochemical and other fields, and it is suitable for the on-site rapid detection, the results of which can be used for real-time on-line analysis and testing the physical and chemical index of the material, otherwise the spectral signal data of the material has a sort of functional relation with its physical and chemical indicators. The current model carrying NIR data mainly includes principal component regression, partial least squares regression, etc. Aiming at the trait of the near infrared spectrum of real-time online material analyzing, and the trait which carries high dimension, less labeled data etc, additionally in the light of the existing incremental support vector regression algorithm, the thesis proposed an incremental semi-supervised ensemble algorithm of support vector regression (IS3VRE).As the traditional supervised learning is applied to the near infrared spectrum data, we usually analyze the labeled spectral data, consequently we missed unlabeled data, which causes the waste of the spectrum data and low generalization ability. Considering that to tag unlabeled data of NIR spectral data is laborious once more, therefore the research of analyzing the near infrared spectrum data by a semi-supervised method has realistic value. Support vector regression machine modeling be used to make the low-dimensional nonlinear input mapped to the high-dimensional linear output. The model is simple and is supposed to have a splendid application prospect. Based on that there is less unlabeled data and more labeled data among the near infrared spectrum data, the thesis is going to adopt incremental semi-supervised ensemble algorithm of support vector regression machine to carry out the research.Based on support vector regression machine as a learning device, the thesis has studied the incremental semi-supervised integrated algorithm of support vector regression machine. First of all, we have constructed an incremental semi-supervised support vector regression model and have chosen the high confidence coefficient data to tag data collaboratively by using the nearest neighbor algorithm, and then we have put forward a method that according to that whether the labeled data can be the potential support vector to decide whether to update the SVR model. Secondly, we have built an incremental semi-supervised ensemble model of support vector regression machine and have put up n support vector regression model, then we have used 10-fold cross-validation deviation to assign weight for each model, and at last have made model integration, in this way could we get an improved regression machine. Finally, having had the model established in this thesis applied to the near infrared spectrum data; we have verified the validity of the algorithm in this thesis and achieved good results by observing the performance between this model and other algorithm model in the case of different data labeling rate. This algorithm applied to the non near infrared spectrum red wine data also achieve good results, which shows that this algorithm can also be applied to other types of data sets.