Collaborative Research: HCC: Medium: Modeling and Mitigating Confirmation Bias in Visual Data Analysis

人们在做出决定时，很容易被数据淹没，比如决定哪种医疗保健治疗合适，或者投票给哪位政治候选人。当被数据淹没时，人们倾向于以一种支持他们先前存在的信念的方式来寻找和解释信息。这种现象通常被称为确认偏差。在数据交流和视觉分析中，确认偏差可能尤其邪恶，即使是对经验丰富的分析师也是如此。尽管有一种误解认为统计模型和可视化提供了客观的事实，但在现实中，在收集、处理、分析和呈现数据时的选择可能会使人们过度依赖他们先前存在的信念。该项目将通过(1)创建模型，显示现有信念和分析目标如何影响数据驱动的决策制定，以及(2)设计新颖的分析界面，通过智能地建议可能反驳信念的证据，帮助分析师做出不那么有偏见的决策，从而仔细检查数据分析中的确认偏差。该项目团队还将创建教育材料并收集经验数据集，以帮助数据分析师、研究人员和公众思考视觉数据交流和解释中的确认偏差。本项目旨在增加对确认偏差在真实世界视觉数据分析任务中如何表现的理解，并开发和评估减轻偏差的干预措施。该项目围绕四个研究项目展开。由于测量确认偏差需要捕获个人的信念，研究人员将首先调查实验方法，以准确捕获一个人在分析环境中关于数据模式和趋势的信念和心理表征(推力I)。然后，研究人员将利用这些方法论的发现来测量低级视觉分析任务中确认偏差的效果并对其进行建模(推力II)，然后在更现实的分析环境中检查这些任务的较高级别的成分(推力III)。最后，研究人员将设计和开发四种缓解偏差的干预措施，将其纳入真实世界的视觉分析工具，如Data Voyager和Jupyter笔记本电脑，并招募专业分析师在数字实地研究中对其进行评估(推力IV)。这项研究议程将促进对确认偏差的理解，并提供有希望的干预措施，使数据分析师能够做出更好的决策。研究人员还将开发课程工作和倡议，将计算机科学、心理学和伦理学结合在一起，以推动围绕视觉数据分析的实践和教育。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 can easily be overwhelmed with data when making decisions, such as deciding which healthcare treatment is appropriate or which political candidate to vote for. When overwhelmed by data, people tend to seek and interpret information in a way that supports their preexisting beliefs. This phenomenon is often referred to as confirmation bias. In data communication and visual analytics, confirmation bias can be especially nefarious, even for experienced analysts. Although there is a misconception that statistical models and visualizations present objective truths, in reality, choices in the collection, handling, analysis, and presentation of data can bias people into overly relying on their pre-existing beliefs. This project will closely examine confirmation bias in data analysis by (1) creating models that show how existing beliefs and analytic goals can impact data-driven decision-making, and (2) designing novel analytic interfaces that help analysts make less biased decisions by intelligently suggesting evidence that may disprove a belief. The project team will also create educational materials and collect empirical datasets to help data analysts, researchers, and members of the public think about confirmation bias in visual data communication and interpretation. This project aims to increase understanding of how confirmation bias manifests in real-world visual data analysis tasks and to develop and evaluate bias-mitigation interventions. The project is structured around four research thrusts. As measuring confirmation bias requires capturing an individual’s beliefs, the researchers will first investigate experimental methods to accurately capture a person's beliefs and mental representations about data patterns and trends in an analytic setting (Thrust I). The researchers will then leverage these methodological findings to measure and model the effect of confirmation bias in low-level visual analytic tasks such as finding correlations (Thrust II), then examine higher-level compositions of these tasks in more realistic analysis settings (Thrust III). Finally, the researchers will design and develop four bias-mitigation interventions to be incorporated into real-world visual analytic tools such as Data Voyager and Jupyter Notebooks, recruiting professional analysts to evaluate them in digital field studies (Thrust IV). This research agenda will advance the understanding of confirmation bias and provide promising interventions to empower data analysts to make better decisions. The researchers will also develop coursework and initiatives that bring together computer science, psychology, and ethics to advance practice and education around visual data analytics.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.