Knockoff Feature Selection Techniques for Robust Inference in Supervised and Unsupervised Learning

该项目旨在开发一种新的方法，用于在预测最终结果的大量潜在变量中选择关键特征。当应用于生物医学领域时，这些方法将能够发现患者健康的决定因素，从而改善疾病的预防、治疗和管理。当在工程、心理学、社会学、经济学和环境科学等领域使用时，这些方法可以改善制造过程、关注多样性和公平性的社会项目、心理健康的护理和管理，以及环境和自然资源的保护。此外，新方法还将有助于生成高质量的合成数据，同时保持原始信息的机密性，从而促进新的科学发现，并提供有价值的教育工具。该项目将在统计学、机器学习和生物医学领域为未来的数据科学家群体提供一系列独特的跨学科培训计划。研究议程基于“仿制方法”，用于识别预测结果的关键特征，同时保持对错误发现的控制。这些方法将微生物组系统发育结构纳入特征选择，适应缺失值，合并多个仿制品以提高鲁棒性，对复杂数据结构采用非参数贝叶斯模型，并引入基于条件预测函数的新的仿制品统计。提出的统计数据可以与最先进的机器学习模型配对，以检测非线性关系，同时考虑特征相关性。此外，通过将仿冒过滤应用于无监督学习模型，本研究可以识别特征空间的决定因素，并为无监督聚类和学习提供见解。该奖项反映了美国国家科学基金会的法定使命，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。

This project aims to develop a new methodology for selecting key features among a large pool of potential variables that are predictive of the final outcomes. When applied to the biomedical field, these methods will enable the discovery of determinants of patient health, thus improving the prevention, treatment, and management of diseases. When used in fields such as engineering, psychology, sociology, economics, and environmental sciences, these methods can improve manufacturing processes, social programs that focus on diversity and equity, the care and management of mental health, and the preservation of the environment and natural resources. Additionally, the new methods will also help to generate high-quality synthetic data while maintaining the confidentiality of the original information, thereby spurring new scientific discoveries and providing a valuable educational tool. The project will offer a number of unique interdisciplinary training initiatives for the future cohorts of data scientists at the interface of statistics, machine learning, and biomedical sciences.The research agenda is based on the 'knockoff method' for identifying key features predictive of the outcomes while maintaining false discovery control. The methods incorporate the microbiome phylogenetic structure in feature selection, accommodate missing values, incorporate multiple knockoffs to increase robustness, employ nonparametric Bayesian models for complex data structures, and introduce a new knockoff statistic based on conditional prediction function. The proposed statistics can be paired with state-of-the-art machine learning models to detect nonlinear relationships while accounting for feature correlation. Furthermore, by applying knockoff filtering with unsupervised learning models, this research can identify determinants of the feature space and provide insights into unsupervised clustering and learning.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.