Mixed-Initiative Visual Text Analytics: Data-driven Views and Analytic Guidance

点击翻译按钮获取中文摘要

Visual analytics is the science of combining data processing, information visualization, and software to support people in the process of data analysis. The strength of visual analytics comes from having computers do work appropriate to their capabilities, such as counting, clustering, and bulk processing of data, while human analysts do the work of hypothesis formation, reasoning, evidence gathering, and decision making. The traditional way to start analyzing data with visualizations is to "overview" the data using high-level visual summaries. However, as data scales have grown, overviews are increasingly challenging to design in an effective way. Overviews of very large data, such as all the Tweets in a city over a year, are often too cluttered to reveal anything interesting. So, rather than starting with an overview, it may be better to suggest a starting point for analysis, such as a single Twitter user, or tweets from a single day. The proposed research will harness computing power to create data-driven suggestions of starting points for analysis and guidance for next steps. There is a trade-off around how far to take the role of computers in analyzing data. On one side, automatic highlighting or hiding of data details introduces potential risk of decisions being biased by algorithms. I argue that burdening analysts with visualizations which do not provide analytic guidance is equally inappropriate, as it ignores the ability of computational systems detect trends and regions of potential interest automatically. Analysts faced with general overviews may not know where to go, and may experience frustration as they start an investigation. By suggesting an interesting place to start, an analyst may more quickly achieve a state of deep engagement in analytic tasks. Thus, this proposal presents a program of research to investigate the nexus of the human-computer relationship in visual analytics: where should the control lie, what sorts of guidance are possible and helpful, and when should guidance be provided? In all of these investigations, I imagine retaining a human-in-charge paradigm, where suggestions and guidance can be ignored or deactivated, similar to the `autocomplete' functions on a mobile phone.  The questions in this research will be investigated using text and document data. Text data is economically and socially important, and generated in enormous volumes daily, through emails, business reports, books, news, legal proceedings, and more. Through the research process, 10 students will receive specialized training in highly sought-after data science skills. This research will introduce new visual analytic practices, achieving a closer coupling of analysts and algorithms.  The outcomes promise to improve Canadians' ability to gain insights from economically and socially important large-scale datasets in domains from business intelligence to health informatics.