Abstract
Graph learning is a critical driver of advances in scientific discovery, business modeling, and AI-assisted decision-making. However, two fundamental roadblocks hinder its broader adoption: the scalability of powerful, subgraph-based learning methods, and the critical need to protect sensitive information within relational data. The complex dependencies inherent in graph structures render many traditional optimization techniques intractable.
This talk introduces a unified, system-aware algorithmic framework to dismantle these scalability and privacy barriers. To address scalability, I will present a novel family of subgraph-based frameworks, which decouple subgraph-level feature extraction from model training and inference by utilizing reduced forms such as random walks, sets, and hashes. This approach has enabled large-scale learning on graphs with billions of edges. The co-design principle is then extended to privacy-preserving graph learning. Here, the structural dependencies in graphs often violate the key assumption of standard privacy mechanisms like DP-SGD and lead to prohibitive memory overheads. I will introduce the first relational learning framework that offers rigorous privacy guarantees. We have successfully deployed this framework to fine-tune large language models on sensitive graph data and achieve a strong privacy utility trade off.
By holistically co-designing learning algorithms and their underlying system implementations, this research demonstrates the efficient and trustworthy application of graph-based AI to real-world problems. This work opens new avenues for leveraging complex structured data, a crucial step towards building the next generation of foundation models .
For more info, please follow this link.