III: Small: Machine Learning on Graphs with Distribution Shifts

One of the key challenges in advancing artificial intelligence is improving AI models' ability to generalize across different tasks, particularly when encountering data distribution shifts from training to real-world applications. Data distribution shifts are encountered when the statistics of the learned model are different from the real-world statistics. This project focuses on enhancing machine learning with graph-structured data, a direction with the potential to revolutionize scientific discoveries in particle physics and biochemistry. In addition to contributing to advancements in these scientific fields, the project also emphasizes knowledge dissemination through curated new datasets, scalable software solutions, workshops, and tutorials. Additionally, it encourages undergraduate and K-12 students from underrepresented groups to engage in research projects through the STEM program at Georgia Tech. This research will tackle the problem of distribution shifts in graph machine learning through two main research thrusts. The first thrust, graph structure calibration, aims to develop methods to estimate and mitigate shifts in entity connection patterns within graph-structured data from training to evaluation phases. It will address tasks such as node classification, regression, and link prediction. Additionally, this thrust will explore the accumulation of graph structure shifts and address challenges in open-world settings. The second thrust focuses on developing foundational models for graph data that are provably expressive, generalizable, and scalable, and investigating robust fine-tuning approaches for these models. The developed methodologies will be evaluated on particle representations in high-energy physics applications and molecule representation tasks. 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.