基于粒计算的多模态多标记数据分类建模研究

The task of multi-label classification learning using multi-modal data description which exists widely in many practical applications such as medical diagnosis, multimedia information retrieval, etc. In multi-modal multi-label data, the multi-modal feature bears the multi-view multi-granularity classification information, and the hierarchical structure of multi-label embeds the multi-granularity structure of classification. It is hence a challenge for existing classification learning approaches to capture the hierarchical heterogeneous multivariate information of the learning task. To fully make use of the structural information of multi-modal multi-label data, the project will research the classification modeling of multi-modal multi-label data based on granular computing. The main contents include: (1) Study the knowledge representation and evolution mechanism of multi-granularity classification pattern in multi-modal feature space; (2) Study the multi-granularity hierarchical structural framework in multi-label decision space, and present a neighbor label classification method using the label correlation; (3) Build the reasoning mechanism of cross-granularity between multi-modal feature space and multi-label decision space, and use the reasoning mechanism to mine the corresponding relation between each modal in the input space and each label in the output space; (4) Based on the drug/disease data, develop new drug reposition algorithms with respect to multi-modal multi-label learning framework. The results of this project will not only enrich the theory of granular computing by providing new solutions for multi-modal multi-label learning, and will also definitely provide new research ideas and methods for the application of drug reposition.

采用多模态数据描述的多标记分类学习任务广泛存在于医疗诊断、多媒体信息检索等领域。其多模态特征承载了多视角多粒度的分类信息、多标记的层次结构嵌入了分类的多粒度结构。现有的分类学习方法难以有效捕获此类任务的层次化异构多元信息。为了充分利用多模态多标记数据的结构信息，本项目基于粒计算思维，研究该类数据的分类建模方法。主要研究内容包括：（1）研究多模态特征空间中多粒度分类模式的知识表示与演化机理；（2）研究多标记决策空间的多粒度层次结构表示框架，并提出利用标记相关性辅助的邻域标记分类方法；（3）构建多模态特征空间与多标记决策空间之间的跨粒度推理模型，用于发掘输入空间中每个模态与输出空间中每个语义标记的对应关系；（4）以药物/疾病数据为载体，设计多模态多标记学习框架下的药物重定位算法。项目研究成果既丰富了粒计算理论，为多模态多标记学习提供新的解决方案，而且为药物重定位应用提供新的研究思路和方法。

多模态多标记数据分类学习建模是国内外机器学习与数据挖掘界的重要研究课题，相关技术广泛应用于药物研发、疾病诊断、视频标注等领域。本项目围绕以多模态多标记学习为代表的多模态多义性对象学习建模技术开展研究，主要研究内容包括：（1）针对多标记高维数据，提出了基于粒计算的多种特征选择与分类学习算法，建立了面向混合数据、流环境的高效特征选择算法；（2）基于信息粒化定义了覆盖度、投票率等多种决策规则度量指标，提出了多模态数据的规则学习方法；（3）利用类属属性设计多模态特征空间与多标记决策空间之间的跨粒度推理模型，构建多种多标记分类学习模型；（4）通过整合药物-疾病两个层面积累的海量多源异构数据，提出了在融合药物多模态信息基础上基于协同过滤算法探索药物的新用途。针对上述研究内容，项目组基于相关研究方案开展工作，完成了原定研究计划和目标。基于本项目成果，共发表学术论文28篇，包括《IEEE Transactions on Fuzzy Systems》、《Pattern Recognition》、《Fuzzy Set and Systems》、《模式识别与人工智能》等国内外重要学术期刊。任国内学术会议特邀报告4次。培养硕士研究生若干。