Study Of Semi-supervised Based On Kernel

Semi-supervised learning has played an important role in pattern recognition and machine learning, which has been concerned by International Machine Learning Session. In recent years, as machine learning has been widely used in data mining and analysis, a number of theories of Semi-supervised learning has been successfully applied into practical issues.Semi-supervised learning has been studied in this paper, whose main content is as follows:Using the generalized kernel consistency method, the semi-supervised learning algorithm named GCM(Generalized Consistency Method) which based on kernel strategy is presented in this paper. Five different measures and the interrelations among them are also deeply analyzed. Relation between arguments of different measures and performance of algorithm is experimentally studied, and performance of GCM algorithm with different measures is compared with each other. Experimental results show that performance of GCM algorithm with the exponential measure is superior to one with other measures and performance of GCM algorithm with the Euclidean measure is inferior to one with other measures. Moreover, some arguments of different measures have a certain effect on the performance of this algorithm. Since CCA is actually a linear learning model, the nature of linear transformation limits to abstract the nonlinear discriminating features of samples, so Semi-CCA dealing with nonlinear problems is inadequacies. We introduce the kernel method in Semi-CCA algorithm and present a semi-supervised learning algorithm based on kernel canonical correlation analysis called Semi-KCCA to transform the nonlinear problems in original space into linear problems in feature space. The performance of this algorithm in dealing with nonlinear problems is studied by experiments. Finally, Consistency Method and its transformation is deeply analyzed and studied; the proof about the important condition for Consistency Method convergence is given in detail; in addition, this paper carries out deep research on the convergence of some other transformations of this algorithm, and proved that using the transformation matrix in CM algorithm is reasonable.