Research On Sparse Bayes Model Applied In Classification And Regression

Bayesian model, which describes the uncertainty relationship between variables by probability, has become the effective modeling method in data mining for its advantages of sound theoretical basis and reasonable inference ability. To alleviate its high time-consuming computation, inspired by Relevance Vector Machine, Sparse Bayesian Model is proposed and the corresponding algorithms for classification and regression is generalized. The main works is following as below:1) Sparse Bayesian Classification(SBC hereafter) is proposed and then applied to the cost-sensitive problem. According to the Automatic Relevance Determinations(ARD), SBC assumes the parameter satisfied the zero-mean Gaussian prior distribution, and then obtains the optimal separating hyperplane by calculating the maximum a posteriori probability with sample information and prior knowledge. Applied the SBC to the cost-sensitive problem and the novel algorithm, Cost-Sensitive Sparse Bayesian Classification(CSSBC) is proposed. Experiments on artificial dataset and UCI datasets show the effectiveness of the CSSBC.2) Sparse Bayesian Regression(SBR) algorithm is proposed and then applied to the novelty detection problem. SBR assumes the noise on regression satisfied zero-mean Gaussian distribution and then calculates the maximum posteriori probability to find the optimal surface. Applied to novelty detection and the new algorithm is proposed. Experiments on artificial dataset and UCI datasets demonstrate the effectiveness of the algorithms.