Pattern mining from linked data, methods, tools and applications

这个程序是关于从发布在Web上的结构化数据中提取有用的知识。它研究了以前贡献的模式挖掘方法和技术对新兴的Web of Data的适应性，Web of Data是一种可在Web上免费获得的链接开放数据(LOD)数据集，可供访问/检索。底层格式是资源描述框架(RDF)，这是一种基于图形的语言。云的LOD数据集通常具有显式的机器可读模式，它们相互链接到一个几乎唯一的全球图表中，该图表正在不断扩展(2011年有320亿条边)。在这样一个巨大的异质但结构相对良好的数据源中进行数据挖掘既是有益的，也是具有挑战性的：(1)LOD数据集的关系结构对于标准数据挖掘者来说是很难处理的，(2)需要适当地利用可用的模式知识来指导挖掘过程。我们的程序专注于模式挖掘，即发现数据描述中的递归规则。所提取的模式在该领域有着广泛的应用，从模式求精到推荐再到决策。然而，现有的模式挖掘器在以下三个方面中至少有一个失败：将图形平面化为特征向量，从而丢失用于挖掘的宝贵信息，忽略模式知识，或者牺牲效率以换取表达能力。我们建议研究RDF数据集的可变模式语言的可伸缩性、表现力和覆盖率之间的更好权衡。这项研究将建立在旨在处理丰富模式(本体)的挖掘方法上，尽管输入数据略有不同且规模较小。因此，关键的挑战是扩大它们的范围和效率。为此，我们将研究如何根据一般数据挖掘问题对它们进行适当的重塑，以便能够应用最先进的数据结构、技术和结构结果。预期的应用是药物不良反应的分析(药物警戒)，在重症监护中的决策，在基因组数据的系统发育分析中设计复杂的工作流程。

This program is about extracting useful knowledge from the structured data that is published on the Web. It researches the adaptation of previously contributed methods and techniques for pattern mining to the emerging Web of Data, a cloud of Linked Open Data (LOD) datasets freely available on the Web for access/retrieval. The underlying format is the Resource Description Framework (RDF), a graph-based language. LOD datasets of the cloud, often with an explicit machine readable schema, are inter-linked into a virtually unique global graph which is constantly expanding (32 billion edges in 2011). Data mining within such a huge heterogeneous yet relatively well structured source is both beneficial and challenging: (1) the relational structure of a LOD dataset is hard to process for standard data miners and (2) the available schema knowledge need to be properly exploited to guide the mining process. Our program focuses on pattern mining, i.e., discovery of recurring regularities in data descriptions. The extracted patterns have a wide range of applications from schema refinement to recommendation to decision making in the domain. However, existing pattern miners fail in at least one of the three ways: flatten the graph into a feature vector and hence lose precious information for mining, ignore the schema knowledge or sacrify efficiency for expressiveness. We propose to research on a better trade-off between scalability, expressiveness and coverage of variable pattern languages for RDF datasets. The research will build upon mining methods designed to work with rich schemas (ontologies) albeit with a somewhat different input data and of smaller size. The key challenge is therefore to increase their scope and efficiency. To that end, we shall research on properly recasting them in terms of general data mining problems to enable the application of state-of-the-art data structures, techniques and structural results. The intended applications are analysis of drug adverse reactions (pharmacovigilance), decision making in critical care, design of complex workflows in phylogenetic analysis of genomic data.