具有不确定性信息的一分类和多分类算法的研究

Support vector machine-based multi-class classification and one-class classification are important topics in the areas of machine learning and pattern recognition. This project aims at resolving the classical research problems in support vector classification and new research challenges arising from the application perspectives. First of all, we design an interactive SVM-based one-against-one algorithm to resolve the unclassifiable regions existing in the original SVM-based one-against-one algorithm, and guarantee the convergence of the developed algorithm. In addition, the developed interactive SVM-based one-against- one algorithm fully utilizes the data information of the unclassifiable regions. Secondly, by analyzing that one can not guarantee the classes of a dataset well linearly separable in a feature space, this project proposes a design tree architecture-based multi-space-mapping algorithm to resolve the problem that the accuracy of the current multi-class classification algorithms is sensitive to the distribution of the dataset. Thirdly, for the problem that the collected data always contain uncertain information due to noise or device accuracy, this project utilizes nearest neighbor idea to measure the bound of uncertainty in the feature space and then incorporates the bound of uncertainty into SVM-based one-class learning phase so that the learned classifier is less sensitive to the uncertainty information. Last, this project proposes a transfer-learning-based one-class SVM algorithm to handle the problem of data distribution transfer in data stream environment. This project aims at developing workable and efficient support vector classification algorithms and delivering insight thoughts to the area of support vector machine.

支持向量机多分类和一分类算法研究是机器学习和模式识别中重要的研究课题。本项目以支持向量机多分类和一分类为研究对象，首先通过构造交互迭代算法解决支持向量机一对一多分类算法的不可分区域问题，并保证算法的收敛性；该算法充分利用不可分区域中的样本信息解决不可分区域问题。其次，通过分析同一特征空间不能保证多分类数据在特征空间线性可分，提出了基于决策树的多特征空间映射算法来解决支持向量机多分类算法的精度对数据分布敏感的问题。同时，针对样本包含不确定信息的问题，本项目通过最近邻思想把不确定信息边界引入支持向量机一分类算法，从而降低算法对不确定信息的敏感性。最后,针对数据分布发生迁移的问题，本项目提出基于迁移学习的支持向量机一分类算法，把迁移前后的数据分布同时引入学习，从而提高分类算法的预测能力。本项目旨在通过以上问题的解决为支持向量机多分类和一分类算法的研究提供思路。

通过本项目的实施，我们提出了嵌套支持向量机来解决一对一多分类中的不可分区域问题，实验结果显示我们的算法在精度方面对多分类算法有显著的提升，并保证了算法的收敛性；其次，我们通过分析同一特征空间不能保证多分类数据在特征空间线性可分，提出了多空间映射算法来解决多分类算法对数据分布敏感的问题，并保证算法的计算复杂度和传统算法保持同一个数量级。针对样本包含不确定信息的问题，我们通过最近邻思想把不确定信息边界引入支持向量机一分类算法，并给出算法的计算复杂度和求解方案，实验验证了算法的有效性。最后, 为了解决数据分布发生迁移的问题，我们提出了迁移学习支持向量机一分类算法，把迁移前后的数据分布同时引入学习，提高分类精度的同时保证算法的收敛。通过本项目，我们为支持向量机多分类和一分类中的以上问题研究提供研究思路和解决方案。