CRII: CHS: Modeling Analysis Behavior to Support Interactive Exploration of Massive Datasets

University Of Maryland, College Park, College Park MD

Investigators

Abstract

Scientists commonly use exploratory data analysis methods to gain insights from their data. However, increases in the number and granularity of data sources raise problems of scale that complicate the already difficult problem of developing tools that help analysts manage the often-changing goals and analysis trajectories suggested by their exploratory work. This project focuses on improving two key systems in exploratory data analysis tools: the visualization systems that provide graphical representations of the data, and the data management systems that efficiently manage large-scale data on the back end to support the analysis. The key idea is to integrate these two systems by first inferring analysts' goals and future behaviors from their recent actions in the visualization system, then using those to proactively construct efficient processing queries in the data management system. Doing this should improve system response times, which should in turn improve analysts' ability to use the system and the insights they gain; the techniques developed will contribute to the database, visualization, and human-computer interaction communities. The tools themselves stand to benefit a number of scientific and industrial domains, and the team will also use the project work to support new interdisciplinary data science courses along with outreach and research opportunities for underrepresented students in computer science. To improve performance, this project will produce dynamic optimization strategies for visual exploration systems, which infer the user's exploratory analysis goals over time, and deploy optimization algorithms tailored to the current analysis goal. These optimizations will address both human performance, i.e., how effectively a scientist or analyst extracts insights and performs analysis tasks with a visual exploration system, and system performance, i.e., how efficiently and effectively the system responds to a user's interactions. The development of these optimizations will be done in two phases. First, a user study will be conducted to characterize how users interact with visual exploration systems under different exploratory data analysis scenarios and system designs. Second, using the collected study data, a predictive query execution engine will be designed to infer users' analysis goals from log data and detect shifts in behaviors over time. To boost data management system performance, existing techniques will be adapted to leverage the predictive query execution engine, including query scheduling of likely upcoming queries and multi-query optimization to leverage computational overlap between recent and predicted queries. To boost visualization system and human performance, the system will recommend predicted next queries to analysts, while the project team will conduct performance-driven interface design work to design new interactions based on data collected by the predictive query execution engine. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.