Abstract: In order to improve the learning performance, semi-supervised learning methods aim at exploiting the knowledge of a small amount of labeled examples as well as lots of unlabeled data instances simultaneously. However, most existing semi-supervised approaches, primarily focus on the effective utilization of those label-unknown data, and the successive study regarding the label-known examples is usually neglected. In light of such situation, in terms of the manifold regularization framework, this paper proposes a novel semi-supervised classification method based on joint regularization of manifold and pairwise constraints (SSC-JRMPC). This method proceeds from two aspects: on one hand, inheriting from the manifold regularization framework, the optimization regarding both empirical risk and structural risk, and the use of intrinsic data distribution of entire dataset are considered concurrently; on the other hand, by transforming the sample labels into the must-link/cannot-link pairwise constraint conditions and incorporating these extended knowledge into own objective formulation, the knowledge existing in the supervision information is further mined. As the results, the classification accuracy of SSC-JRMPC is distinctly enhanced. The experiments on real-world datasets confirm the merits of this paper work.

Key words: semi-supervised learning, classification, manifold regularization, pairwise constraints

摘要： 半监督学习方法主要通过学习少量标记样本和大量未标记样本知识来提高学习效果，然而目前许多半监督方法注重在未标记样本的利用上深耕，忽略了对标记样本等监督信息的继续研究。鉴于此，结合流形正则化框架提出了一种流形与成对约束联合正则化半监督分类方法（semi-supervised classification method based on joint regularization of manifold and pairwise constraints，SSC-JRMPC）。SSC-JRMPC从两个方面进行研究：一方面该方法继承了流形正则化框架中的特点，将经验风险和结构风险最小化，以及对整个数据的内在数据分布进行运用；另一方面，通过将样本标签转化为成对约束的形式，并把这些扩展的知识并入到目标公式中来进一步探索监督信息包含的知识，一定程度上提高了SSC-JRMPC算法的分类准确性。通过在真实数据集上的实验，验证了上述优点。

关键词: 半监督学习, 分类, 流形正则化, 成对约束