基于类属特征学习的高效鲁棒多标记学习方法研究

In Multi-label learning, each object may belong to one or multiple class labels simultaneously, and label-specific feature learning is one of the main research methods for multi-label learning. The challenging problems of multi-label learning under big data environment, e.g., high dimension, multiple heterogeneous forms, missing data and dynamic evolution, put great challenges to multi-label learning tasks. In this project, we seek to improve the performance of multi-label classification by building efficient and robust multi-label learning models based on label-specific feature learning. First, to improve the accuracy and robustness of multi-label learning models, we build multi-label learning algorithms by learning label-specific features for each class label and study the sharing mechanism of label-specific features between different class labels by exploiting label correlations. Second, based on label-specific feature learning, we aim to construct pattern representation and fusion mechanism for multiple heterogeneous data. Then we make the discriminant analysis for each single view data representation and the label-specific features of it, and learn the multi-label multi-modal fusion mechanism by learning label-specific data representation. Last, to solve the missing and dynamic changes of class labels, we aim to learn label correlations and build multi-label learning models with missing labels, and propose new multi-label learning models to discover and depict emerging class labels based on label-specific feature learning. The researching results of this project will provide the theoretical basis and technology solutions for multiple Internet applications, e.g., document categorization, image, and video annotation, social network, commodity recommendation.

多标记学习中每个对象同时属于一个或多个类别标记，类属特征学习是当前多标记学习的主要研究方法之一。大数据环境下多标记数据具有海量高维、多源异构、数据缺失和动态变化等特性，使得多标记分类难度更大。本项目采用类属特征学习方法，研究高效鲁棒的多标记学习算法以提升多标记分类任务的性能。首先，针对海量高维性，学习对每个类别具有强判别力的类属特征集合，研究类属特征的共享机制，降低模型复杂度，提升模型准确性和鲁棒性。其次，在类属特征学习的基础上构建多源异构数据的模式表达与融合机制，学习单源数据表示及其类属特征的判别能力，研究基于类属特征的多标记多模态融合机制。最后，结合类属特征学习提出新类别标记发现及其语义描述方法，研究标记缺失下的标记相关性学习以及分类模型的建立方法，以适应标记缺失和动态变化。本项目研究成果将为文档分类、图像/视频标注、社交网络、商品推荐和信息检索等互联网应用提供理论依据和技术解决方案。

针对大数据环境下多标记数据的海量高维、多源异构、数据缺失和动态变化等挑战难题，本项目展开了一系列的相关研究。在本项目的执行过程中发表学术论文9篇，其中SCI/EI检索7篇，申报国家发明专利11项，培养研究生10名，已毕业5名。.本项目在执行过程中较好的完成了预期研究目标，主要研究内容与成果包括三个方面：（1）提出了基于类属特征学习、多标记标记缺失学习、基于图嵌入和特征降维的多标记学习方法，提升多标记分类模型的学习效率和鲁棒性；（2）提出了一种基于视角类属特征学习的多视角多标记学习算法以及基于“视角-标记”贡献度和一致性的多视角多标记分类方法，降低了模型的空间和时间复杂度，提升了模型准确性；（3）提出了基于类属特征学习的多标记学习中未知标记发现和分类方法，实现未知标记的发现、预测以及语义描述。.通过本项目的研究，在一定程度上解决了多标记数据的海量高维、多源异构、数据缺失和动态变化等难题。本项目研究成果将为多标记学习在实际应用中提供理论依据和技术解决方案。