Planning: DCL EPSCOR: CISE Large: Investigations in Foundational Models for Recommendation with Applications to Complex Domains

Recommender systems use machine learning to learn about people's preferences from their feedback. These systems then suggest a small, relevant subset of options from a large pool, tailored to each user's needs. Recommender systems help users find items they might not have discovered on their own, changing how they consume information and reducing the time spent searching. Despite significant progress over the years, recommender systems still face challenges with fairness, including issues of bias and transparency. On the other hand, generative AI, in particular Large Language Models (LLMs), are creating significant new opportunities as well as potential challenges for information access. There is a crucial need to find better, sustainable ways to continue innovating with LLMs while ensuring more equitable benefits. This requires working on fundamental issues of fairness, robustness and trustworthiness and, more generally, a responsible use of AI with respect to communities and the environment. For all these reasons, Large Language Models, when used to power recommender systems, and especially when used to expand human exploration, can have a significant impact on society, in terms of education, economic impact, and equity. This planning award will support the team to build on promising preliminary work and ideas to mount a competitive large-scale impactful CISE core project and proposal, driven by ambitious research, education, and broadening participation goals. The team will investigate novel algorithms for incorporating foundational models, in particular large language models in recommendation systems while anchoring all methods in expanded fairness criteria. The research will address core issues of fairness, robustness, and trustworthiness, and, more generally, a responsible use of AI with respect to communities and the environment in which they live. In particular the research will address multifaceted fairness desiderata that transcend traditional model predictive performance and fairness metrics throughout the various stages of the Large Language Model-based recommender system pipeline. During the planning phase, the team will engage with diverse stakeholders to assess the needs and shape their ideas. In addition to having a number of novel ideas with promising preliminary work, the team has identified several potential application domains with different use cases. These ideas and use cases need to be narrowed down and organized within a competitive and sound plan, along with the formation of an interdisciplinary team from among a potentially large number of stakeholders throughout the team’s institutions and communities. The planning grant will allow the team to incorporate several of their preliminary research ideas int use-inspired research and further mount a comprehensive plan with direct impact on education and broadening participation. 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.