Excellence in Research: Mitigating Confounding Errors in Real-World Machine Learning for Robust Decision Support

Alabama State University, Montgomery AL

Investigators

Abstract

User-generated contents, such as online reviews or social media posts, often contain hidden information like education, personal preferences, location, or language. This information, known as confounding factors, can affect the contents and impact the outcomes of decision systems. When applying machine learning for real-world decision support, those confounding factors can easily have negative effects on model generalizability and usability. This project focuses on identifying and mitigating confounding errors in real-world machine learning to ensure decision systems in healthcare or e-commerce produce more accurate and robust support. The project will generate broad impacts through regional research collaboration, AI workforce development, and practical applications. It will establish a partnership between Alabama State University and the University of Memphis, while insights from the research will be integrated into new machine learning and AI courses at Alabama State University to equip students with advanced technical skills. Additionally, students from all groups will gain experience with large-scale Artificial Intelligence (AI) and Graphic Processing Unit (GPU) computing through the high performance computing cluster hosted by the University of Memphis. Collaborations with industry partners such as FedEx will further translate research into practical, real-world applications. All activities will be open to participants from all groups, ensuring inclusive engagement. The research will develop a contrastive learning framework to uncover and integrate meaningful patterns in these confounding factors, enhancing the robustness and reliability of machine learning models. The project consists of three main efforts: (1) building machine learning models to identify key confounding factors from textual and non-textual data; (2) enhancing these models through iterative learning that incorporates multi-source and multimodal data, such as social media posts and reviews; and (3) employing multi-view learning to mitigate confounding errors and strengthen model reliability. Together, these efforts will establish a robust framework for real-world decision support systems. 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.