Ridge Regression And Active Learning Based On Margin Distribution

Support Vector Machine (SVM) has been one of the most famous algorithm in machine learning. The basic idea is to maximize the minimum margin. However, recent studies show that margin distribution rather than the minimum margin is crucial to performance of learning algorithm. Based on the theory of optimal margin distribution, this thesis makes two studies as follows:1) Propose the ODMRR approach based on the principle of optimal margin distribution for ridge regression by considering the local interclass relationship, as well as the global relationship from different classes. This approach tends to improve the generalization performance of learning algorithm and has closed-form solution and can be easily applied for kernel. Extensive experiments are conducted to verify the effectiveness of the proposed algorithm.2) Propose the ODMAL approach for active learning based on the principle of optimal margin distribution by considering the informative and representative instances. to improve the efficiency of active learning. This approach tends to improve the generalization performance of active learning and has closed-form solution and can be easily applied for kernel. This thesis verifies the proposed algorithm by experiment.