离异图消歧引导的偏标记学习

偏标记学习是一种弱监督学习框架,它试图从样本的多个候选标签中选择唯一正确的标签.消歧是偏标记学习中的一种重要手段,主要通过算法判别潜在的真实标签.目前,人们普遍采用单一的特征空间或者标签空间进行消歧,容易导致算法受到不准确先验知识的引导而陷入鞍点.针对消歧过程中特征相似样本易受到异类样本影响从而影响消歧效果这一问题,定义了样本离异点和离异图;在此基础上,提出一种离异图引导消歧的偏标记学习方法.该方法利用标签空间的差异构建离异图,可以有效结合特征空间的相似性和标签空间的差异性,降低离异点为消歧过程带来的潜在风险.实验结果表明,与PLKNN、IPAL、SURE、PL-AGGD、SDIM、PL-BLC、PRODEN等方法相比较,所提出的算法在偏标签学习方法中表现更好,能够取得良好的消歧效果.

Introducing outlier graph to disambiguate for partial label learning

Partial label learning is a weakly supervised learning framework, which attempts to select the only correct label from multiple candidate labels of the sample. Disambiguation is an important means in partial tag learning, which mainly uses algorithms to identify potential real tags. At present, researchers generally use a single feature space or label space for disambiguation, which is easy to lead the algorithm to fall into saddle point under the guidance of inaccurate prior knowledge. In order to solve the problem that similar samples are easy to be affected by abnormal samples in the process of disambiguation, sample outlier and outlier graph are defined. On this basis, a partial marker learning method for dissociation graph guided disambiguation is proposed. The algorithm uses the difference of label space to construct the divorced point map, which can effectively combine the prior knowledge of feature space and label space, and restrict each other to resist the potential high risk brought by divorced samples to the disambiguation process. Experimental results show that compared with PLKNN, IPAL, SURE, PL-AGGD, SDIM, PL-BLC and PRODEN, the proposed algorithm performs better in the partial label learning method and achieves good disambiguation effect.