EAGER: Collaborative Assembly of Large and Comprehensive Causal Networks

人们强烈希望找到解释和解决方案来解决我们国家和世界面临的紧迫问题，但他们往往缺乏适当的分析工具，无法超越有限范围的分析、猜测和直觉。这个项目旨在通过开发一套工具来解决这个问题，这些工具可以利用网络上免费获得的海量数据。任何拥有现代网络浏览器的人都将能够运行这些工具，并参与合作努力，为复杂的社会经济和其他系统构建全面的因果模型。揭示多变量数据集中变量之间的因果关系是数据分析的最终目标之一。一个因果断言区分了因果关系，例如，它说“吸烟致癌”，但不是相反。这就是因果模型比相关性更明确的原因。因果模型之所以吸引人，是因为它们本身就是可解释的。它们能够直接解释底层数据中存在的复杂交互作用。在这个项目中开发的在线工具将把因果建模提高到一个新的水平。它们将支持合作，为复杂的社会经济和其他系统构建规模空前的综合因果模型。这一探索性项目旨在开发一套工具，利用免费获得的网络规模的数据，为复杂的社会经济和其他系统协作构建规模空前的综合因果模型。该项目将在如何利用专家和非专家的创造性能量来(1)识别网络上可以为不断演变的因果模型添加新方面(变量)的数据集，以及(2)将这些新方面(变量)作为新节点和因果边缘整合到模型中方面开辟新的天地。由于随着模型规模的增大，构建复杂的大规模因果模型可能会变得困难，因此该项目将产生几个新的自动化工具，这些工具将对人类用户隐藏这种复杂性，帮助他们处理不完整或不利的数据，并为可能的改进提供灵感。同时，还将开发新的技术，以确保在并发用户存在的情况下演变的因果模型的有效性和正确性。为了隐藏复杂性，该项目将产生新的技术，可以将一个大型因果模型分解成一组人类可以管理的子图，但这些子图仍将保留有关需要改进的特定主题方面的足够信息。子图将以因果流程图的形式可视化，该流程图可以有效地显示因果关系的传播，并支持可能缺乏足够的领域知识、直觉或其他有用信息的用户，以确定可能使模型更具表现力的有希望的变量。该项目将开发基于单词嵌入范例的新技术，以帮助用户进行这一发现过程。单词嵌入将大型文本语料库中类似上下文中提到的单词映射到高维空间中的紧密邻域中。将开发一种类似于2D地图的可视化，其将在模型中已经存在的语义相关变量的标签附近的因果子图的主题上下文中映射单词(表示候选变量)。然后，人体模型编辑者可以检查这个词的视觉地图(候选变量)，从这些新变量假设可能的新因果关系，在网络上或在演变的因果模型中搜索相关数据，并使用系统的因果推理引擎测试新的因果关系并将其嵌入到子图中。在幕后，一个自动化的因果网络管理器随后会将因果边派生到其他变量，从而全面发展模型。由于在观察数据存在的情况下自动因果推理有时会产生错误方向或无方向的边缘，该界面还将提供新的范例，允许人类模型评估员验证生成的边缘并提出更改建议。将进行一系列精心设计的用户实验，以验证和优化所有系统组件。这项研究有望在以人为中心的计算因果推理和利用网上可用的大量数据方面产生新的理论知识和算法。它还将提供关于人类如何与推导和探索因果模型的工具交互的新见解。在这项研究中产生的平台和工具将适用于多个知识领域，并能够构建能够解释不同方面和领域在更大背景下如何相关的因果模型。开发的工具将作为在线平台的一部分提供。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

People are highly motivated to find explanations and solutions to address the pressing problems facing our country and our world, but they often lack the proper analytical tools to move beyond limited-scope analyses, guesswork, and instincts. This project aims to tackle this problem by developing a set of tools that make use of the massive amount of data that is freely available on the web. Anyone with a modern web browser will be able to run these tools and take part in a collaborative effort to construct comprehensive causal models for complex socio-economic and other systems. Uncovering the causal relations that exist among the variables in multivariate datasets is one of the ultimate goals in data analytics. A causal assertion separates cause and effect, for example, it states that "smoking causes cancer", but not the reverse. This is what makes causal models more definite than correlation. Causal models are attractive since they are inherently interpretable. They are able to directly explain the complex interactions that exist in the underlying data. The online tools developed in this project will take causal modeling to the next level. They will support collaboration in constructing comprehensive causal models of unprecedented scale for complex socio-economic and other systems. This exploratory project aims to develop a set of tools that use freely available, web-scale data for the collaborative construction of comprehensive causal models of unprecedented scale for complex socio-economic and other systems. The project will break new grounds on how the creative energies of experts and non-experts can be harnessed to (1) identify datasets on the web that can add novel aspects (variables) to an evolving causal model, and (2) integrate these novel aspects (variables) as new nodes and causal edges into the model. Since building a complex, large-scale causal model can become difficult as the model grows in size, the project will produce several new automated tools that will hide this complexity from the human users, aid them in dealing with incomplete or adverse data, and provide inspiration for possible refinements. At the same time, novel techniques will also be developed that ensure validity and correctness of the evolving causal model in the presence of concurrent users. In order to hide complexity, the project will produce new techniques that can break a large causal model into a set of human-manageable subgraphs which will nevertheless retain sufficient information about the particular thematic aspect to be refined. A subgraph will be visualized in the form of a causal flowchart that can effectively show the propagation of causal relationships, and support users who may lack sufficient domain knowledge, intuition, or other helpful information to identify promising variables that could make the model more expressive. The project will develop new techniques based on the paradigm of word embeddings to assist users in this discovery process. Word embeddings map words mentioned in similar contexts in large text corpora into close neighborhoods in high-dimensional space. A 2D map-like visualization will be developed that maps words (denoting candidate variables) in the causal subgraph's thematic context near the labels of semantically related variables already in the model. Human model editors can then inspect this visual map of words (candidate variables), hypothesize possible new causal relations from these new variables, search for associated data on the web or in the evolving causal model, and test and embed the new causal relations into the subgraph using the system's causal inference engine. Behind the scenes, an automated causal network manager will then derive causal edges to other variables and so fully evolve the model. Since automated causal inference in the presence of observed data can occasionally generate wrongly directed or undirected edges, the interface will also provide new paradigms that allow human model appraisers to verify the generated edges and suggest changes. A set of carefully designed user experiments will be conducted to verify and optimize all system components. The research is expected to yield new theoretical knowledge and algorithms on human centered computational causal reasoning and the utilization of the vast body of data available online. It will also deliver new insights on how humans interact with the tools for deriving and exploring causal models. The platform and tools generated in this research will be applicable to multiple fields of knowledge and enable construction of causal models capable of explaining how the various aspects and fields relate in a larger context. The developed tools will be made available as part of the online platform.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.