多标签降维中的多重代价敏感学习问题研究

To deal with the high-dimensional data in multi-label learning, multi-label dimensionality reduction attracts researchers’ attention. Multi-label dimensionality reduction mainly concerns the label correlations, however the label correlations is difficult to be obtained correctly, and may easily influenced by multiple cost sensitive problem, i.e., in the label set there is noisy annotation, class imbalance problem in the single label, and inherent cost sensitive problem between labels. Therefore, this subject researches the multiple cost sensitive problems in multi-label dimensionality reduction by adopting the cost sensitive learning methods. More specifically,.1)Assuming the label correlations can be obtained, by the graph learning theory, this subject embeds the multiple costs in label covariance matrix and hypergraph, which are used for representing pairwise label correlations and high-order label correlations respectively, and researches the cost sensitive correlations; Furthermore, to learn the label correlations, this subject proposes the joint learning model of multi-label dimensionality reduction and cost sensitive correlations, which can learn and use cost sensitive label correlations effectively..2)By fusing the methods mentioned above, this subject researches the collaborative learning model of multiple costs and cost sensitive correlations, which can solve the multiple costs learning problem. .This subject will publish more than 6 papers, including more than 2 journal papers published in or above CCF B, more than 2 international conference papers published in or above CCF C, more than 2 papers published in domestic journal; apply for or authorized 1 invention patents and 1 software copyright; assists cultivating 3 master students.

为解决多标签学习中的高维数据问题，多标签降维引起了学者关注。多标签降维主要关注标签相关性，但其难准确获取，且受多重代价敏感问题影响，即标签集上存在噪声标注、单个标签上的类别不平衡以及标签间内在的代价敏感问题。为此，本课题采用代价敏感学习方法，研究多标签降维中的多重代价敏感问题。具体如下:.1）假设标签相关性可得，对表示成对标签相关性的标签协方差矩阵、高阶标签相关性的超图，采用图学习理论，嵌入多重代价，研究代价敏感的相关性；进一步，研究标签相关性的学习，提出多标签降维和代价敏感相关性的联合学习模型，有效学习、利用相关性。.2）协同上述方法，研究“多重代价的学习+代价敏感的相关性学习”的协同学习，解决多重代价的学习问题。.本课题将发表6篇以上论文，包括CCF B类以上期刊和C类以上会议各2篇以上，国内一级学报2篇以上；申请或授权发明专利和软件著作权各1项；协助培养3名硕士研究生。

本项目研究多标签降维中的多重代价敏感问题，即标签集上存在噪声标注、单个标签上的类别不平衡以及标签间内在的代价敏感问题。课题组从构建鲁棒、判别的代价敏感多标签学习模型和多重代价的学习问题这两个研究目标出发，主要研究如下四个方面的研究内容：（1）借助样本特征空间辅助构建与学习标签相关性，获得鲁棒、判别的标签相关性；（2）基于矩阵分解技术的代价敏感样本特征提取，用于代价敏感多标签降维模型的构建；（3）采用两步以及联合学习策略的半监督代价敏感学习模型，以获得准确的样本标签信息，进而定义样本的错分代价信息；（4）扩展研究面向多标签数据的隐私保护问题。课题的研究成果可应用于图像分类、视频自动标注、情感分类、信息检索等诸多实际应用领域。.受项目资助，共发表（录用）11篇学术论文，包括CCF A类期刊2篇，B类期刊1篇，C类期刊3篇，EI收录的国际会议论文3篇，其中一篇获得IEEE BigMM'2018国际会议最佳论文提名，2篇中文核心论文；申请发明专利1项；培养以及协助培养硕士研究生4名。