Research Classification System Of Large Celestial Spectra Data Based On Support Vector Machines

In recent years, there have been very significant developments in the theoretical understanding of Support Vector Machines (SVMs) as well as algorithmic strategies for implementing them, and applications of the approach to practical problem. The thesis researches the classification methods based on Support Vector Machines by taking the national science engineering project LAMOST as background. Hierarchical Support Vector Machines based on clustering proposed implements multi-class classification about stellar spectra Data. The main works are as follow:At first, integrating chunking algorithm into Sequential Minimal Optimization (SMO) algorithm, the thesis proposes dynamic chunking SMO algorithm. Using thought of chunking algorithms, training samples set is divided into two subsets, one as working set, SMO executed on it, according to the optimization result, some samples are selected to compose new working set from another subset. Experimental results show that the algorithm can reduce further the training time.Secondly, the thesis proposes a new classification method called hierarchical Clustering Support Vector Machines (HC-SVMs) on the basis of researches of multi-class SVMs proposed. The method decomposes training samples set using clustering analysis, i. e. clusters the similar class, and constructs the hierarchical structure, then trains the classifier. The classification method uses decomposition strategy to decompose initial problem into a series of two-class classification problems that reduces training scale and enhances training speed. And, ultima multi-class classifier includes a few sub-classifiers, and therefore enhances classification speed.Finally, Support Vector Machines classification system on stellar spectra data faced to LAMOST project is designed and realized using VC++ and Oracle9i, and its software architecture and function modules are outlined. The preliminary experimental results show the system is workable for stellar spectra data classification.