异构大数据的对象建模及跨域分析技术研究

This research project deals with the challenging problems of cross-domain big data, i.e., heterogeneous forms, complicated correlation and dynamic evolution. It aims at building the object modeling, cross-domain learning and system feedback mechanisms, and realizes never-ending knowledge acquisition from cross-domain heterogeneous big data. The proposed cross-domain heterogeneous big data analytics are of great value for the national scientific developing strategies on big data and World-Wide-Web, and satisfy the needs of social information management. This project plans to construct the cross-domain object modeling methods and cross-domain analytics from aspects of data, model, system and application, respectively. The research objectives are as follows. First, we aim to obtain the primary relation between data items and objects via consistent representation learning, intrinsic representation extraction and category label extraction. Second, we aim to construct structurally consistent cross-domain object linkage functionalities by combining the pattern representation learning and attribute analysis. Third, based on pattern representation learning and object modeling in the first two steps, we aim to construct the cross-domain knowledge transfer and temporal knowledge transfer mechanisms. Fourth, based on object information and cross-domain transfer learning, we propose the correlation computation method between objects, and study methodologies for object network construction, cross-domain object-based multiple instance learning and collaborative correlation prediction. Finally, we aim to construct the knowledge representation mechanism based on object network, and provide solutions for personalized recommendation on cross-domain heterogeneous big data.

本课题针对跨域大数据的异构形式、复杂关联、动态演化等挑战性问题，研究对象建模、跨域学习和系统反馈等机制，实现永不终止的跨域异构大数据知识获取。所提出的跨域异构大数据分析技术框架，对满足国家大数据和互联网科技战略以及社会信息化管理等需求具有重要价值。本课题拟从四个层面：数据、模型、系统和应用来构建跨数据域的对象建模和跨域分析方法，包括提取异构数据的一致性描述、本征表示、类别标签等信息，得到其与对象的初步关联；结合异构数据的模式表达和属性分析，建立结构一致性的跨数据域对象链接机制；基于跨域模式表达和对象建模，建立横向知识迁移（跨域迁移）和纵向知识迁移（时序迁移）机制；基于对象结构信息和跨域差异分析，提出对象之间的关联计算方法，构建跨域异构大数据对象网络，探索跨域多实例学习，实现异质关联协同预测；基于异构对象学习，构建知识表达机制，为跨域异构大数据的个性化推荐系统提供技术支持。

本项目旨在建立以高维数据域、图数据域和关系数据域为代表的异构大数据的对象建模、跨域学习和反馈处理机制，实现永不终止的跨域异构大数据知识获取系统过程。在项目执行期间，以跨媒体作为跨域异构大数据的代表性数据形态，围绕研究目标开展了深入的研究工作。在异构数据的统一表征、对象提取与关系计算、跨域学习等方面取得关键技术突破：提出了跨媒体调和学习与深度统一表征等理论方法框架，实现了跨媒体数据的拓扑结构保持学习和多层次信息的对齐协同表征；针对不同模态数据内容的特点，提出了面向不同模态，多粒度异构时空数据对象的精确检测和关系度量计算机制，并在此基础上实现了不同模态之间的复杂内容转换计算，为大规模跨媒体知识图谱构建提供了关键技术支撑；针对跨媒体内容的复杂时空演化和知识匮乏滞后等难题，提出了开放域迁移、批量数据输出核范数最大化、渐进式/启发式迁移等多种方法，实现了面向开放域的跨域异构知识迁移与复用；并进一步在组件、机理、过程、结果等方面实现了不同层面的可解释性分析推理。基于已取得的研究成果，构建了人机协同的跨媒体知识图谱标注系统，面向多个跨媒体应用场景开发了原型验证系统。共发表(及录用)期刊和会议论文37篇，其中IEEE/ ACM汇刊论文13篇，CCF-A类顶级国际学术会议论文20篇(大会宣讲论文5篇)，申请发明专利4项，授权1项，累计(协助)培养毕业博士生3人，硕士生6人。