Planning Grant: I/UCRC for Computation Intensive Big Data Analytics for Multimodal Temporal Prediction, Retrieval, and Attribution

点击翻译按钮获取中文摘要

Recently, both leading technological firms and academic institutions have developed a new capability to capture and process data at a scale many orders larger than previously possible. Firms, particularly in Silicon Valley, have demonstrated how this capability, when blended with Analytics, can be exploited to solve a host of new problems of great practical value. These range from Computational (or Online) Advertising, with immense commercial value being realized, to Healthcare Analytics, which is of tremendous and critical value societally. At our Research Site at the University of California, Santa Cruz, we focus on enhancing the predictions and searches based on Variety of data (in addition to Volume and Velocity), since data such as Electronic Healthcare Records (EHRs) include not just numerical data (about vitals and labs), but also text (notations by doctors and nurses), images (X-rays etc.), video (such as a doctor?s examination), body sensors, etc. We also explore and exploit ways of identifying more informative data on-the-go, as well as identifying effective ways to speed up and slow down the rate of obtaining this informative data automatically based on need and context, to achieve superior prediction. Finally, we develop new ways of evaluating the true impact of each data type/source on the desired outcome. The fundamental discoveries will concern methods for determining the value of each type or source of data in more effective prediction (and search) of dynamic system state and intervention decisions.Our research concerns analyzing multi-type and source data for enhanced prediction, search, and decision making. Our collaborative research site focuses on research and development of novel scalable Big Data Analytics (BDA) solutions for multimodal data and streaming data suitable for various computing architectures (where the knowledge extracted from multiple modes is far greater than the sum of knowledge discovered from individual data-types); and novel interoperable solutions to combine analytics tools and solutions. Specific sub-thrusts include:1. Scalable temporal prediction/classification, including recommenders, for multi-type data, incorporating latent business processes and ontologies.2. Bayesian interactive information retrieval for massive data sets, combined with extraction incorporating latent business processes and ontologies in mining data.3. Data type and source characterization in terms of power of prediction/classification/retrieval, including causality in marketplaces and temporal aspects.4. Interoperable solutions for BDA that enable seamless and efficient knowledge discovery from multimodal data involving different solutions and tools in different languages and different APIs, and Programmable File and Storage Systems for Big Data Analytics.5. User behavioral modeling analytics and the analytics/economics of decisions and interventions.The domains and contexts we propose to explore include a subset of: system health (e.g. aviation safety, jet engine maintenance) and analytic services based on heterogeneous data and data/text mining, extraction, and retrieval for service centers (e.g. in network health or financial, sales and marketing services), principled knowledge discovery for personalized healthcare and web analytics, diagnostics and prognostics in Internet of Things.