CAREER: Large-Scale Multi-label Learning

科学和商业中的重要应用依赖于将大量项目(示例)自动分类到预定义的类别中。多标签学习是指每个样本可以同时被分配到多个类别标签的分类问题。分类问题涉及到许多不同的领域，如自然语言处理、计算机视觉、人机交互、生物信息学、医疗保健和生理学。现有的机器学习技术不适合大规模多标签学习，因为它们不能处理罕见的类分类问题，并且不能很好地区分具有相似输入模式的类。为了克服现有方法的局限性，该项目正在开发一个多标签学习的关系传播框架，该框架显式地同时利用示例的相似性和类之间的相关性。具体而言，本项目的研究包括：(1)开发高效的关系传播框架优化算法；(2)开发有效的样本相似度和类间相关性学习算法；(3)开发有效的多标签学习主动学习算法；(4)通过文本分类、图像标注和基因表达模式预测这三个实际应用对所提出的多标签学习框架进行评估。该项目将通过开发关系传播框架来促进大规模多标签学习技术的发展，这将对广泛的应用产生重大影响。研究成果还将加强目前的机器学习课程，让学生参与跨学科研究，并改善信息技术工作人员的教育。该项目的网站(http://www.cse.msu.edu/~rongjin/))将用于传播研究成果。

Important applications in science and business depend on automatic classification of large volumes of items (examples) into predefined classes. Multi-label learning refers to the classification problem where each example can be assigned to multiple class labels simultaneously. The classification problem is germane to many different domains, such as natural language processing, computer vision, human computer interaction, bioinformatics, health care, and physiology. Existing machine learning technologies are unsuitable for large-scale multi-label learning because they are unable to handle rare class classification problems and poorly distinguish classes with similar input patterns. To overcome the limitation of the existing approaches, the project is developing a relation propagation framework for multi-label learning that explicitly exploits the similarity of examples and the correlation among classes simultaneously. In particular, this project includes research to (1) develop efficient optimization algorithms for the proposed relation propagation framework; (2) develop effective algorithms for learning the similarity of examples and the correlation among classes; (3) develop effective active learning algorithms for multi-label learning; and (4) evaluate the proposed framework for multi-label learning through three real world applications: text categorization, image annotation, and prediction of gene expression patterns.The project will advance the state of the art of techniques for large-scale multi-label learning through the development of relation propagation framework, which in return will have a significant impact on a wide range of applications. The research results will also enhance the current machine learning curricula, involve students in interdisciplinary research and improve the education of the information technology workforce. The project Web site (http://www.cse.msu.edu/~rongjin/) will be used for research results dissemination.