Information Theoretic Research on Big Data Compression and Analytics: Theory, Algorithms, and Applications

随着我们社会中数据集的爆炸式增长，我们正处于大数据革命的开端。大数据有可能加快科学、工程和医学领域的发现步伐，改善医疗保健、金融、商业和我们的生活，并最终改变我们的社会。 然而，为了利用大数据革命带来的机遇，必须认真应对大数据带来的许多挑战。例如，为了快速远程访问，并降低维护大量数字数据的持续成本，期望尽可能多地压缩数据。同样，为了从大量的数字数据中收集知识和发现，必须开发基于上下文分析的工具和技术来组织、可视化和管理大量的数字数据。高效的数据压缩和准确的上下文分析对于将大数据转化为知识、将发现转化为行动，同时实现资源（带宽和电力）效率至关重要。 在过去，数据压缩和数据分析在很大程度上是在不同的学科中单独研究的。虽然数据压缩已经存在多年，但其现有的理论和技术，特别是来自信息论的源编码理论，主要是针对静态和/或结构良好的数据（如文本，网页等）而开发的。然而，在大数据的背景下，大多数数据类型是非固定的和非结构化/半结构化的;将现有的压缩技术直接应用于这些数据类型通常会导致压缩性能不满意。因此，必须开发新的压缩理论和技术，将数据分析纳入数据压缩中，以优雅的方式处理结构和平稳性的缺乏。相反，研究如何将信息理论思想，特别是源代码理论和技术应用于数据分析，以开发更好的数据分析解决方案，如认知距离，聚类和组织，也是有益的。 在我们早期成功的基础上，在这项研究中，我们将调查和探索数据压缩和分析的相互作用，以促进这两个领域的知识。我们的方法将是信息理论。三个主题领域将集中在：（1）数据压缩和分析的相互作用，以提高非静态和非结构化数据的有损压缩性能;（2）数据压缩和分析的相互作用，以开发更好的认知距离，从而为数据聚类和组织提供更好的工具;（3）大型二分图的压缩和模式发现。在实现科学目标的过程中，我们将保持和加强在相关领域的领导地位，并为加拿大培训信息和通信技术方面的高素质人才。

With an explosion in data sets in our society, we are at the beginning of a big data revolution. Big data has a potential to accelerate the pace of discovery in science, engineering, and medicine, improve healthcare, finance, business, and our lives, and ultimately transform our society. To tap the opportunities afforded by the big data revolution, however, many challenges of big data have to be carefully addressed. For example, to fast access remotely, and reduce the on-going cost of maintaining, huge volumes of digital data, it is desirable to compress data as much as possible. Likewise, to glean knowledge and discoveries from huge volumes of digital data, tools and techniques based on context analytics have to be developed to organize, visualize, and manage huge volumes of digital data. Efficient data compression and accurate context analytics are of vital importance to turning big data to knowledge and discoveries to actions while achieving resource (bandwidth and power) efficiency. In the past, data compression and data analytics have been largely investigated separately in different disciplines. Although data compression has been around for many years, its existing theories and techniques, especially source coding theory from information theory, have been largely developed for stationary and/or well-structured data (such as text, web pages, etc.). In the context of big data, however, most data types are nonstationary and unstructured/semi-structured; applying existing compression techniques directly to these data types often leads to unsatisfied compression performance. Therefore, it is imperative to develop novel compression theories and techniques by incorporating data analytics into data compression to handle the lack of structure and stationarity in an elegant way. In the opposite direction, it is also beneficial to investigate how to apply information theoretic ideas, particularly source coding theory and techniques, to data analytics to develop better solutions for data analytics such as cognitive distance, clustering, and organization. Building on our early success, in this research, we will investigate and explore the interactions of data compression and analytics to advance knowledge in both fields. Our approach will be information theoretic. Three theme areas will be focused on: (1) interactions of data compression and analytics to improve lossy compression performance for nonstationary and unstructured data; (2) interactions of data compression and analytics to develop better cognitive distances and hence provide better tools for data clustering and organization; and (3) compression of and pattern discovery in large bipartite graphs. In the process of achieving our scientific objectives, we will maintain and enhance our leadership in the related areas, and train highly qualified personnel in information and communications technology for Canada.