CAREER: Modeling Dependencies via Graphs: Scalable Inference Methods for Massive Datasets

职业：通过图建模依赖关系：海量数据集的可扩展推理方法

• 批准号: 1254106
• 负责人: Animashree Anandkumar
• 金额: $ 56.06万
• 依托单位: University of California-Irvine
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2013
• 资助国家: 美国
• 起止时间: 2013-02-01 至 2019-01-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1254106&HistoricalAwards=false
• 关键词: CAREER; Modeling; Dependencies; via; Graphs

This research centers on theoretical and applied research on learning and representation of high-dimensional data. The term high dimensionality refers to the property that the number of variables or unknowns is typically much larger than the number of observations available at hand. A key challenge is being able to represent and learn such phenomena with sample and computational requirements scaling favorably in the number of dimensions. This project addresses these challenges through a graphical approach by exploiting the inherent graphical structure present in many large data-sets.This research considers modeling high-dimensional data through probabilistic graphical models, also known as Markov random fields. An important research thrust of this proposal is to develop novel algorithms for learning and inference under the framework of graphical models. Another important thrust of this proposal is to develop efficient scalable models for representing high-dimensional data beyond the traditional framework of graphical models. This research establishes strong theoretical guarantees for the developed methods, as well as applies them to real data in various domains, including genetic and financial data, and data from large online social networks such as Facebook and Twitter.

本研究集中于高维数据学习和表示的理论和应用研究。术语“高维”是指变量或未知数的数量通常远大于现有观测值的数量。一个关键的挑战是能够在样本和计算要求在维数上有利地扩展的情况下表示和学习此类现象。该项目通过利用许多大型数据集中存在的固有图形结构，通过图形方法解决这些挑战。该研究考虑通过概率图形模型（也称为马尔可夫随机场）对高维数据进行建模。该提案的一个重要研究主旨是在图模型框架下开发新的学习和推理算法。该提案的另一个重要推动力是开发高效的可扩展模型，用于表示超出传统图形模型框架的高维数据。这项研究为所开发的方法建立了强有力的理论保证，并将其应用于各个领域的真实数据，包括遗传和金融数据，以及来自 Facebook 和 Twitter 等大型在线社交网络的数据。