一般多视图核机的监督和半监督学习方法的研究

In this project, the theories and methods of supervised and semi-supervised learning for general multi-view kernel machines are studied systematically and thoroughly. Based on consistency and complementary information between views, supervised learning methods for general multi-view kernel machines are designed. The contents include general multi-view support vector machines, general multi-view twin support vector machines and general multi-view maximum entropy discrimination based on soft margin consistency. The semi-supervised learning methods of general multi-view kernel machines are designed. The contents include general multi-view semi-supervised least squares support vector machine with consensus manifold regularization and general multi-view semi-supervised least squares twin support vector machine with consensus manifold regularization, in order to make breakthroughs in the design of general multi-view kernel machine classifier. General multi-view multi-label twin support vector machines are studied. General multi-view twin support vector machines are used to deal with multi-view multi-label data. Multi-view deep twin support vector machines are studied which integrate multi-view twin support vector machines and deep neural networks to achieve great success in dealing with large-scale multi-view data. Various theoretical properties of general multi-view kernel machines, such as the characteristics of solutions of optimization problems, sparsity and generalization performance of the classifier are analyzed mathematically and statistically in order to better enhance the understanding of the classifiers and it can provide an important guiding scheme for solving practical problems.

本项目对一般多视图核机的监督和半监督学习的理论与方法展开系统深入的研究，以视图间的一致性和互补信息为切入点，设计一般多视图核机的监督学习方法，内容包括一般多视图支持向量机、一般多视图双平面支持向量机和基于软间隔一致性的一般多视图最大熵判别。并设计一般多视图核机的半监督学习方法，内容包括一般多视图一致性流形规范化半监督最小二乘支持向量机和一般多视图一致性流形规范化半监督最小二乘双平面支持向量机，力求在一般多视图核机的分类器设计方面取得突破性进展；对多视图多标签双平面支持向量机进行研究，将一般多视图双平面支持向量机处理多视图多标签数据；对多视图深度双平面支持向量机进行研究，融合多视图双平面支持向量机和深度神经网络的模型，力求在处理大规模多视图数据取得巨大成功；对一般多视图核机的各种理论性质如优化问题解的特性、分类器的稀疏性和泛化性能等进行数学与统计分析，以更好地增强对它的认识，并对实际应用问题

多视图学习越来越受到广泛的关注，本项目对一般多视图核机的监督和半监督学习的理论与方法展开系统深入的研究，以视图间的一致性和互补信息为切入点，设计一般多视图核机的监督和半监督学习方法，内容包括多视图半监督支持向量机，一般多视图广义特征值近似支持向量机，多视图K近似平面聚类，多模态步态识别算法，非平行平面支持向量机的主动学习和多视图深度双平面支持向量机。通过对这些内容进行理论研究与实验验证，提出了合理有效的算法。本项目的研究对于多视图学习的理论研究和实际应用具有重要意义。