基于多源异构数据关联学习的自我认知技术研究

Self-awareness is quite important for intelligent computing systems. The data of computing systems collected from multiple sources are usually heterogeneous, related and dynamically changing. Current works mostly focus on single factor or cope with a specific task while building self-aware models. Systematic research regarding various factors of self-aware computing systems is rarely seen. This project focuses on the study of relational learning on multi-heterogeneous graphs for self-aware computing system, and conducts the research from the aspects of fusion of multi-heterogeneous graphs, relational learning on multi-heterogeneous graphs and temporal migration in relational learning. First, we will focus on the low-loss and low-noise methods of fusing multi-heterogeneous graphs, which give a unified expression of multi-heterogeneous graphs and therefore enhance the information of self-awareness. Second, we will investigate how to abstract the requirements of self-aware computing systems and construct ubiquitous models of relational learning. We will focus on learning the relations between structures and attributes, the relations between heterogeneous entities and the relations between composite entities of different granularity, and realize self-awareness based on the learning results subsequently. Third, we will build entity-entity-temporal tensors and their corresponding latent factor models to make predictions across large temporal scales or real time predictions, which will make self-awareness techniques more practical. All research will be conducted and tested on a cluster rendering system. The project is expected to yield innovative outcomes of self-aware techniques, as well as theoretical and practical methods of relational learning on multi-heterogeneous graphs.

自我认知技术是计算系统智能化的关键技术。计算系统数据有多源异构、关联耦合和动态变化等特点，而现有工作大多只针对某一因素或某一具体任务构建静态认知模型，缺少系统性研究。本项目聚焦基于多源异构数据关联学习的自我认识技术，从多源异构数据融合、多源异构数据关联学习和动态认知关联学习三个方面展开研究。首先研究多源异构数据融合方法，目标是实现低损低噪融合和统一表达，建立认知信息基础；其次研究如何抽象自我认知问题并为其建立普适的关联学习模式，重点进行结构和属性之间、异质实体之间、高层大粒度复合实体之间的关联学习，目标是辨析它们之间的相互作用和因果关联，建立认知方法；最后研究动态认知关联学习，构建实体—实体—时间张量及其隐特征分析模型，目标是实现大时间跨度和实时性关联预测，使认知技术有实用性。本项目在一个集群渲染系统上展开试验，不仅将创新自我认知技术，亦将产生兼具理论和实用价值的多源异构数据关联学习方法。

数字技术在各个领域的广泛应用，从各种来源积累了大量的结构化、异质性、关联耦合和动态变化的数据，认知这些复杂数据对于推动人类的生产活动和科学发展具有重要意义。本项目研究多源异构数据之间的关联学习方法，研究工作围绕以下内容展开：（1）复杂异构图数据的关联分析方法；（2）多源异构图数据的融合方法；（3）异构图数据在隐特征向量空间上的表达学习方法；（4）多源异构图数据张量融合及张量关联学习方法；（5）异构图数据表达学习和关联发现的图神经网络模型。.项目取得的主要研究成果包括：（1）建立了异构图数据的关联分析模型，实现了图的结构和节点属性之间，以及异构图数据上变化粒度实体之间的关联分析；（2）提出了基于图乘和张量的多源异构图数据融合方法，实现了多源异构图数据无损融合；（3）提出了属性异构图和动态异构图在隐特征向量空间上表达学习的方法，建立了变化粒度基元表达学习的统一计算模式；（4）提出了多源异构图数据张量融合方法和张量上非对称关联关系的学习模型，有效利用多源耦合信息，解决了高阶张量的稀疏问题；（5）建立了异构图表达学习和关联发现的图神经网络模型。提出了图神经网络消息增强传播机制，解决了过平滑和过拟合问题。设计了学习模型的注意机制和多通道神经网络融合模型，实现了动态异构图的关联预测。.本项目的研究成果为融合多源异构图数据进行关联分析提供了系统性的方法框架和实现技术，为学习和认知不同应用领域的复杂高维关联结构数据开拓了新途径。在项目支持下，在高水平学术期刊和国际会议上发表了14篇学术论文，培养了十余位硕士研究生。..