Pattern Discovery and Analysis for Mixed-Mode Data and Sequences: Theory and Applications

拟议的研究将为混合模式数据和序列的模式发现和分析开发有效和可扩展的方法，重点是理论发展、在工业和生物信息学中的应用以及为学术界和工业界培养高素质的人才。 在静态数据库中，混合模式数据采用连续值和离散值的混合值，其中离散值可以是序数，也可以是名义值。在动态过程中，每个数据点，无论是连续的还是离散值的，都与一个时间戳相关联。因此，该采集的数据是连续时间序列和称为混合模式序列的离散特征序列的混合。随着无线传感器的发展，这些类型的数据在今天非常常见，并将继续扩大，这对现有的数据挖掘技术提出了重大挑战，特别是在石油、能源、金融和生物信息学等应用中。 为了应对这一挑战，拟议的研究扩展了我的大学团队和衍生公司多年来所做的研究，以推进对非常大的混合模式数据和序列的模式发现和分析。我们提出的方法已经从将模式定义为整个空间中的元组到子空间中的统计事件关联；(2)从大量发现的模式到总结的模式簇；(3)从事件关联到概率结构表示(例如属性和随机超图)；(4)从从数据本身发现模式到将元数据紧密地结合到发现过程中。为了很好地扩展到超大型数据库，需要开发有效的剪枝技术、高效的数据结构和可扩展的算法。超图模型揭示了数据中固有的微妙知识，并使领域专家能够做出明智的决策--这是在这个PB时代朝着“从数据到知识”的挑战迈出的重要一步。在该提案中，将重点培训HQP，培养他们在理论和应用方面的实力，以及技术转让和商业化技能。

The proposed research will develop effective and scalable methodologies for pattern discovery and analysis of mixed-mode data and sequences with emphasis on theoretical development, applications in industry and bioinformatics, and high quality personnel training for academia and industry. In a static database, mixed-mode data takes on a mixture of continuous and discrete values, where the later can be either ordinal or nominal. In a dynamic process, each data point, be it continuous or discrete valued, is associated with a time stamp. This acquired data then is a mix of continuous time series and discrete character sequences referred to as mixed-mode sequences. With the advancement of wireless sensors, these types of data are very common today and will continue to expand, posing a significant challenge to existing data mining techniques especially for the applications such as oil, energy, finance and bioinformatics. To respond to this challenge, the proposed research extends many years of research done by my university team and spinoff companies to advance pattern discovery and analysis for very large mixed-mode data and sequences. Our proposed approaches have moved (1) from defining patterns as tuples in the entire space to statistical event associations in subspaces; (2) from overwhelming number of discovered patterns to summarized pattern clusters; (3) from event associations to probabilistic structural representations (such as attributed and random hypergraphs); (4) from discovering patterns from the data itself to incorporating metadata closely into the discovery process. To scale well to very large databases, effective pruning techniques, efficient data structures and scalable algorithms need to be developed. The hypergraph models uncover subtle knowledge inherent in data and enables domain experts to make informed decisions - an important step towards the challenge "from data to knowledge" in this petabyte era. In the proposal, strong emphasis will be placed on the training of HQP in developing their strength in theory and applications as well as in technology transfer and commercialization skill.