III-COR: Collaborative Research: Mining Biomedical and Network Data Using Tensors

IIS 0705359，IIS 0705215 III-COR：Collaborative Research：Mining Biomedical and Network Data Using Tensors Christians Faloutsos（christos@cs.cmu.edu）CMU Vasileios Megalooikonomou（vasilis@cis.temple.edu）Temple Univ.鉴于随着时间的推移功能磁共振（fMR）图像的大量收集，如何找到模式和相关性？同样，如果网络流量信息源源不断，如何监控异常、入侵和潜在故障？这个提议背后的主要思想是使用张量理论来处理这两个问题。尽管这两种设置看起来有很大的差异，但它们都归结为在多维数组中寻找模式，稀疏或密集。张量是矩阵的推广，大致对应于数据挖掘的“DataCubes”。 矩阵分析和分解是数据挖掘标准工具箱的一部分，提供了降维、模式发现和"隐藏变量“发现的方法。将这些工具扩展到更高的维度是有价值的，张量提供了进行这种推广的工具。然而，这些工具尚未投入使用，在大容量的数据挖掘。这是本提案的主要贡献。研究人员建议（a）设计张量分解算法，可扩展到大型数据集，特别注意稀疏数据集，以及永无止境的数据流和（B）将它们应用于两个驱动应用程序，fMRI数据分析和网络数据分析。研究人员建议分析大量的功能磁共振成像数据，执行以下子任务：聚类体素与相似的行为随着时间的推移为一个给定的主题和/或任务或跨主题和/或任务，分类模式的大脑活动，并检测滞后相关性和时空模式之间的功能磁共振成像时间序列。研究人员还建议对多千兆位的网络流数据执行以下相互关联的任务：异常检测，模式发现和压缩。这两个应用程序对于医学，健康管理以及计算机和国家安全都很重要。对功能磁共振成像数据的分析可以帮助理解大脑的功能，大脑的哪些部分与其他部分合作，以及是否存在跨学科和跨任务相关活动的差异。对于网络流量监控设置，快速检测异常非常重要，可以发现恶意软件、端口扫描尝试以及普通的非恶意故障。教育目标包括将研究结果纳入CMU（15-826）和Temple的高级研究生课程（9664，9665），并在数据库、数据挖掘和生物信息学领域的领先会议上提供教程。http://knight.cis.temple.edu/~vasilis/research/tensors.html

IIS 0705359, IIS 0705215III-COR: Collaborative Research: Mining Biomedical and Network Data Using Tensors Christos Faloutsos (christos@cs.cmu.edu) CMU Vasileios Megalooikonomou (vasilis@cis.temple.edu) Temple Univ.Given a large collection of functional Magnetic Resonance (fMR) images over time,how can one find patterns and correlations? Similarly, given a never-ending stream of network traffic information, how can one monitor for anomalies, intrusions, and potential failures? The main idea behind this proposal is to treat both problems using the theory of tensors. Despite the seemingly wide differences in the two settings, they both boil down to finding patterns in multidimensional arrays, sparse or dense. Tensors are exactly generalizations of matrices, and correspond roughly to ``DataCubes'' of data mining. Matrix analysis and decompositions are part of the standard toolbox for data mining, providing methods for dimensionality reduction, pattern discovery and``hidden variable'' discovery. Extending these tools to higher dimensionalities is valuable and tensors provide the tools to do this generalization. However, these tools have not yet been put to use in large volume data mining. This is the main contribution of this proposal. The investigators propose (a) to design tensor decomposition algorithms that scale for large datasets,with special attention to sparse datasets, and to never-ending streams of data and (b) to apply them on two driving applications, fMRI data analysis and networkdata analysis. The investigators propose to analyze large volumes of fMRI data performingthe following sub-tasks: cluster voxels with similar behavior over time fora given subject and/or task or across subjects and/or tasks, classify patterns of brain activity, and detect lag correlationsand spatio-temporal patterns among fMRI time sequences. The investigators also propose to perform the following inter-related tasks on multiple GigaBytes of network flow data: anomaly detection, pattern discovery, and compression.Both of these applications are important for medicine, health management,and for computer and national security. Analysis of fMRI data can help understandinghow the brain functions, which parts of the brain collaborate with what other parts, and whether there are variations across subjects and across task-related activities. For the network traffic monitoring setting, fast detection of anomalies is important,to spot malware, port-scanning attempts, and just plain non-malicious failures.The educational goals include incorporating the research findings in advanced graduate courses at CMU (15-826) and at Temple (9664, 9665)and proposing tutorials in leading conferences in databases, data mining and bio-informatics audiences.For further information see the web page: http://knight.cis.temple.edu/~vasilis/research/tensors.html