Bayesian Modeling and Scalable Inference for Big Data Streams

我们正处于一场数据革命之中。在数据测量、收集和生成技术的最新进步的推动下，数据的激增正在重塑众多科学和应用学科的格局。基于互联网的公司必须通过使用大量可用的用户数据来为个人量身定做推荐和广告来争夺消费者的注意力；金融交易公司必须每天从新闻、公司报告和市场中合成数TB的数据来进行知情投资；微生物学家现在面临着分析数以万计的单个细胞的整个转录组的问题；还有更多的问题。数据的爆炸式增长也引发了对“数据分析师”的重新定义：不再只是那些受过深度统计和数学培训的分析师，随着他们的主要挑战转向处理各自领域的大规模流媒体数据，越来越多的分析师来自其他技术学科。这给计算机科学和统计学的交叉点带来了挑战：我们需要从数据中学习的算法和模型，这些算法和模型在计算上容易处理，能够跟上不断涌现的大规模流数据的步伐；但为了得到从业者的信任，它们还必须易于实施和使用，并对学习结果的质量提供严格的理论保证。我的研究的基本目标是通过为现代大规模和流数据开发有效、实用和易于使用的概率机器学习方法来应对这些挑战。我的研究提案涉及一种应对大数据挑战的多方面方法，主要有三个方面的贡献：1)易于使用的、理论上合理的大规模数据推理算法2)流数据的灵活模型和推理算法3)学习模型和近似的质量的理论分析本研究计划的发展将通过开放源代码发布向更广泛的社区提供，总体目标是使统计建模可供日益多样化的专业从业者使用，并适用于日益多样化的大规模、流数据分析问题。

We are in the midst of a data revolution. Driven by recent advances in data measurement, collection, and generation technologies, the proliferation of data is reshaping the landscape of numerous scientific and applied disciplines. Internet-based companies must vie for consumer attention by using the vast quantities of available user data to tailor recommendations and advertisements to individuals; financial trading firms must synthesize terabytes of data daily from the news, company reports, and markets to make informed investments; microbiologists are now faced with analyzing the entire transcriptome of tens of thousands of individual cells; the list goes on. This explosion in data has also caused the redefinition of "data analyst:" no longer just those with in-depth statistical and mathematical training, analysts are arising more and more from other technological disciplines as their main challenges shift towards dealing with the large-scale, streaming data in their respective fields. This presents challenges lying at the intersection of computer science and statistics: we need algorithms and models for learning from data that are computationally tractable and can keep pace with the constant deluge of large-scale, streaming data; but to be trusted by practitioners, they must also be easy to implement and use, and come with rigorous theoretical guarantees on the quality of the learned result. The fundamental goal of my research is to address these challenges by developing effective, practical, and easy-to-use probabilistic machine learning methods for modern large-scale and streaming data. My research proposal involves a multifaceted approach to the challenges of big data, with contributions in three main areas: 1) Easy-to-use, theoretically sound algorithms for inference with large-scale data 2) Flexible models and inference algorithms for streaming data 3) Theoretical analysis of the quality of learned models and approximations The developments in this research program will be made available to the broader community through open-source code releases, guided by the overarching goal of making statistical modeling accessible to the growing diversity of practitioners and applicable to the growing diversity of large-scale, streaming data analysis problems.