Smart visualizations of multivariate knowledge

威廉·普莱费尔(1759-1823)被许多人认为是统计图形之父，他很可能会被现代消费者电子表格软件迷住。在一个程序中，用户可以放置数千条数据，只需点击几下鼠标，就可以计算汇总值，并以饼图、条形图、折线图和散点图的形式显示数据的分布。此外，对单个数据点的更改可以通过统计数据立即传播到图形，让外行人员可以轻松地生成图形。但他也可能会惊讶于他一生中开发的许多可视化技术，以及目前的软件通常一次只能处理几个变量之间的关系。我认为，我们正处于研究前沿的一个点上，在这个点上，可以从观测数据中实时自动生成关于大量变量之间的概率和因果关系的复杂猜测，以直观的方式将这种关系可视化，识别人类生成的猜测(其中计算机生成的猜测是模棱两可或不完整的)中的弱点和优势，并有效地使用这些图表来理解政策变化的潜在后果。本工作借鉴并扩展了我以前的工作以及其他人在人工智能、图形模型、数据可视化、信息可视化和计算机图形学领域的工作成果，以产生算法和技术，这些算法和技术将为普通数据用户(即未受过培训的统计学家)提供直观的可视界面，用于有效地探索数据集的推测空间，并在必要时进行交互。已经存在软件实现的概念验证，并且已经发布了初步结果。

William Playfair (1759-1823), considered by many to be the parent of statistical graphics, would likely have been dazzled by modern consumer spreadsheet software. In a single program, one can place thousands of pieces of data, and with a few mouse clicks, compute summary values, and display distributions of data in pie, bar, line, and scatter graphs. Furthermore, a change to a single data point can be propagated immediately through the statistics and to the graphics, letting lay persons easily produce the graphs. But he also might have been surprised at how many of the visualizations were ones he developed in his lifetime, and at the fact that current software generally manipulates relationships between only a few variables at a time. I believe we are at a point in the research frontier where it is possible, in real time, to automatically generate complex conjectures about probabilistic and causal relationships among large numbers of variables from observational data, to visualize such relationships in an intuitive way, to identify weaknesses and strengths in human generated conjectures (where computer-generated conjectures are ambiguous or incomplete), and to use these diagrams effectively to understand the potential consequences of policy changes. The present work draws on and extend results from my previous work, and the work of others, in the areas of artificial intelligence, graphical models, data visualization, information visualization, and computer graphics to produce algorithms and techniques that will give ordinary users of data (i.e, persons who are not trained statisticians) an intuitive visual interface for productively exploring the conjecture space of datasets, interacting as necessary. A proof of concept software implmentation already exists, and preliminary results have already been published.