Hierarchical Machine Learning for Information Networks

人工智能（AI）机器学习是未来的一项关键技术，加拿大政府和企业对此进行了大量投资。最先进的人工智能系统使用最丰富的数据。特别有价值的是描述对象之间链接的信息网络。这些链接可以是物质的，如计算机网络中的链接，也可以代表抽象的关系，如社交网络中的友谊，或图像中的空间关系。信息网络无处不在，由许多组织在关系数据库中维护。我的研究项目的目标是开发新的机器学习方法，利用、分析和集成信息网络中异构且相互依赖的数据源。信息网络的机器学习在多项数据分析任务中实现了最先进的性能，例如基于链接的分类和聚类、链接预测、数据库查询优化、异常挖掘、异常和欺诈检测。 尖端应用程序正在从包含文本和/或图像的海量数据集中提取网络信息。例如，谷歌研究人员将机器学习应用于网络信息，构建了一个包含 5.7 亿个节点（实体）和 18 亿个事实（关系和属性）的信息网络（称为知识图谱）。 拟议的研究开发了利用类本体来学习大型信息网络的图形模型的方法。图模型（例如贝叶斯网络或因果图）可以被视为概率知识库，它表示节点、链接以及节点和链接的属性之间的统计模式。因此，图形模型学习提供了大型知识库的自动构建，支持许多数据分析任务，包括基于链接的预测和信息提取。在许多复杂的领域中，本体或类层次结构可以提供有关领域结构的宝贵知识。例如，在一所公立大学中，教员是部门雇员，部门雇员是大学雇员，政府雇员是政府雇员。我的研究将调查类层次结构如何提高统计结论的有效性、统计模型的准确性以及信息网络统计学习的计算可扩展性。 信息网络机器学习是下一代人工智能应用的基础技术，例如从海量文本和视觉数据中提取结构化信息。所提出的研究将贡献重要的组成技术、图形模型学习和自动化知识库构建。学生将接受培训，开发信息网络机器学习方法，并将其应用到加拿大工业中。

Machine learning for Artificial Intelligence (AI) is a key technology for the future, which has seen strong investment from Canadian governments and businesses. The most advanced AI systems use the most informative data. Especially valuable are information networks that describe links between objects. These links can be material, as in a computer network, or represent abstract relationships, like friendships in a social network, or spatial relationships in an image. Information networks are ubiquitous, maintained by many organizations in a relational database. The goal of my research program is to develop new machine learning methods that leverage, analyze, and integrate the heterogeneous and interdependent data sources in an information network. Machine learning for information networks has led to state-of-the-art performance in several data analysis tasks, such as link-based classification and clustering, link prediction, database query optimization, exception mining, anomaly and fraud detection. A cutting-edge application is extracting network information from massive data sets containing text and/or images. For instance, Google researchers have applied machine learning to web information to build an information network (called the knowledge graph) that contains 570M nodes (entities) and 1.8 billion facts (relationships and attributes). The proposed research develops methods that leverage a class ontology to learn a graphical model for a large information network. A graphical model (such as a Bayesian network or a causal graph) can be viewed as a probabilistic knowledge base that represents statistical patterns among nodes, links, and attributes of nodes and links. Graphical model learning therefore provides automated construction of large knowledge bases, which support many data analysis tasks, including link-based predictions and information extraction. In many complex domains, ontologies or class hierarchies are available that provide valuable knowledge about the structure of a domain. For example, in a public university, a faculty member is a department employee, who is a university employee, who is a government employee. My research will investigate how class hierarchies increase the validity of statistical conclusions, the accuracy of statistical models, and the computational scalability of statistical learning for information networks. Machine learning for information networks is a fundamental technology for next-generation AI applications, such as extracting structured information from massive text and visual data. The proposed research will contribute an important component technology, graphical model learning and automated knowledge base construction. Students will be trained to develop methods for machine learning from information networks, and to apply them in Canadian industry.