Toward Automated Uncertainty Quantification in Causal Inference

在研究因果关系或根据数据做出重要决策时，研究人员和决策者经常会遇到不确定性，这些不确定性可能会影响他们结论的可靠性和可信度。理解和量化这些不确定性对于做出明智的选择至关重要，无论是在科学实验、政策制定还是设计机器学习系统中。为此，该项目旨在开发能够有效解决因果推理中不确定性量化的挑战的算法。特别是，目前许多量化因果关系不确定性的方法依赖于复杂的数学技术，这些技术可能与现实世界的情景不太相符。这个项目试图通过设计既在理论上合理又在实践中适用于广泛情况的算法来改变这种情况。该项目还为研究生提供了研究培训机会。从技术上讲，该项目将侧重于由两个来源引起的因果推断的不确定性：不确定的因果关系图和现有数据的有限信息量，这两个来源都对因果结论和下游决策具有重大影响。为了应对这些挑战，该项目将开发算法，提供灵活的和统计上有效的不确定性估计，最大限度地减少对特定因果问题的依赖。利用算法稳定性和私有数据分析技术的最新进展，将构建因果估计的可信区间，即使因果关系图是不确定的或从数据中学习的。此外，这些置信度区间将与可用的领域知识相结合，以进一步量化由于有限的领域知识和数据的识别能力而产生的不确定性。总而言之，该项目将促进不同不确定性的整合，最终导致因果推理中更可靠和自动化的不确定性量化。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。

When studying cause-and-effect relationships or making important decisions based on data, researchers and decision-makers often encounter uncertainties that can impact the reliability and trustworthiness of their conclusions. Understanding and quantifying these uncertainties is crucial for making informed choices, whether in scientific experiments, policy-making, or designing machine learning systems. To this end, the project aims to develop algorithms that can effectively address the challenge of uncertainty quantification in causal inference. In particular, many current approaches to quantifying uncertainty in causal relationships rely on sophisticated mathematical techniques that may not align well with real-world scenarios. This project seeks to change the situation by designing algorithms that are both theoretically sound and practically applicable across a wide range of situations. This project also provides research training opportunities for graduate students. Technically, the project will focus on uncertainties in causal inference arising from two sources: uncertain causal graphs and limited informativeness of available data, both of which have significant implications for causal conclusions and downstream decision-making. To tackle these challenges, this project will develop algorithms that provide flexible and statistically valid uncertainty estimates, minimizing their dependence on specific causal problems. Leveraging recent advancements in algorithmic stability and private data analysis techniques, confidence intervals for causal estimates will be constructed, even when the causal graph is uncertain or learned from data. Additionally, these confidence intervals will be integrated with the available domain knowledge to further quantify the uncertainty arising from limited domain knowledge and the identification power of the data. Taken together, this project will facilitate the integration of different uncertainties, ultimately leading to more reliable and automated uncertainty quantification in causal inference.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.