CRII: III: Dynamic Prompting and Pruning for Measuring and Controlling Memorization in Text-Attributed Graphs

Artificial intelligence (AI) models are increasingly used to process large-scale structured data, yet their ability to memorize and regurgitate information presents both opportunities and risks. This project advances science and technology by developing methods to measure and control memorization in text-attributed graphs (TAGs), which are widely used in social networks, citation networks, and biological systems. While memorization in machine learning can enhance recall of frequently used information, it also raises concerns related to privacy and security. Existing research lacks a comprehensive framework for evaluating and regulating memorization in text-attributed graphs, particularly concerning how graph structures influence memorization patterns. This project will establish new techniques for assessing memorization, develop methods for dynamically adjusting memorization levels, and create the first benchmark for studying memorization effects in text-attributed graph-based learning. These innovations will strengthen the reliability, privacy, and interpretability of AI models used in critical applications such as healthcare, cybersecurity, and knowledge discovery. By ensuring that AI models can retain useful information while preventing unintended leaks of sensitive data, this work will contribute to the development and deployment of responsible AI technologies. This project develops a rigorous framework for measuring and controlling memorization in TAGs through three key research thrusts. The first thrust introduces a novel dynamic prompting strategy that adapts to input variability, enabling more precise measurement of memorization rates. The second thrust proposes a new dynamic pruning framework that allows for fine-grained control over memorization, ensuring that models can be optimized for either enhanced recall or increased privacy. The third thrust establishes a benchmark for memorization in TAGs, systematically evaluating how memorization is influenced by graph topology, such as node connectivity and long-distance relationships. The project’s methodologies integrate insights from graph neural networks and large language models, bridging gaps in understanding between structured and unstructured data representations. By addressing fundamental challenges in memorization, this research will provide practical tools and insights that benefit AI developers, regulatory bodies, and industries that rely on trustworthy machine learning models. The findings will be disseminated through open-source tools, benchmark datasets, and academic collaborations, fostering broader impact in the AI research community. 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.