CAREER: From Fragile to Fortified: Harnessing Causal Reasoning for Trustworthy Machine Learning with Unreliable Data

尽管机器学习（ML）方法很受欢迎并取得了成功，但它们仍然存在很大的局限性。该项目解决了由于对模型开发所用数据的要求不切实际而产生的限制。具体来说，ML方法需要大量无错误和准确的数据，而这些数据并不总是可用的。在这个项目中，开发了通过利用因果推理处理不完美数据的ML方法。该项目的技术贡献将使临床医生获得的因果知识与不完善的数据相结合，以解决慢性疼痛，这是一个影响数百万人的关键健康问题，每年花费数十亿美元。除了对管理慢性疼痛的影响外，该项目还将通过与研究目标相关的项目实现本科生和研究生教育。该项目有三个重点。Thrust 1将利用已知的因果机制来创建强大而有效的预测模型。与目前通过限制其平滑度来惩罚模型容量的方法不同，该项目将重点关注惩罚模型编码概念与已知因果机制相矛盾的方案。重点2将通过建立因果敏感性分析来开发量化预测中的不确定性的方法，并开发量化非参数独立性检验中的不确定性的方法。Thrust 3将利用已知的因果机制来创建强大而高效的离线强化学习（RL）模型。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Despite their popularity and success, machine learning (ML) methods still have major limitations. This project addresses limitations that arise because of unrealistic requirements concerning the data used for model development. Specifically, ML methods require massive amounts of data that is error-free and accurate, which is not always available. In this project, ML methods that work with imperfect data by leveraging causal reasoning are developed. The technical contributions made in this project will enable combining clinician-acquired causal knowledge with imperfect data to tackle chronic pain, a critical health issue impacting millions and costing billions annually. Beyond its impact on managing chronic pain, this project will enable undergraduate and graduate student education through projects related to the research objectives. The project has three thrusts. Thrust 1 will leverage known causal mechanisms to create robust and efficient predictive models. Unlike current approaches that penalize model capacity by constraining its smoothness, this project will focus on schemes that penalize models encoding concepts that contradict known causal mechanisms. Thrust 2 will develop methods to quantify uncertainty in predictions by building upon causal sensitivity analysis, and developing methods to quantify uncertainty in nonparametric independence tests. Thrust 3 will leverage known causal mechanisms to create robust and efficient offline reinforcement learning (RL) models. It will also enable credible estimation of uncertainty to guide policy optimization.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.