Zijie Huang

Research Abstract:

Our world consists of vast amounts of dynamical systems in scientific domains, ranging from brain networks, physical simulations, social networks, to epidemic networks, and have a wide application in different disciplines such as physics, biology, social science, material science, and public health. Many of these can be represented as a graph, where nodes or agents interact with each other and co-evolve governed by certain dynamical laws. By modeling these complex systems, we can gain a deeper understanding of their underlying mechanisms, make more accurate and long-term predictions, and ultimately make well-informed decisions. In line with this, my research vision is to develop deep graph learning models for scientific discovery (Graph ML + Science), spanning from discovering unknown dynamic laws from observational data to accelerating scientific simulations. Towards this goal, my past research has pioneered a simple yet effective framework called Graph Ordinary Differential Equation (GraphODE) which combines the expressive power of data-driven Graph Neural Networks (GNNs) with the symbolic knowledge preserved by ODEs. Compared with most existing discrete methods, GraphODEs show superior performance in long-range predictions and can handle irregular and partial observations. Building upon this backbone, I further made significant advances to improve its expressive power, generalization and causal reasoning abilities. My future research aims to realize my vision via three primary thrusts. Firstly, I aim to develop a scalable and general-purpose machine learning pipeline to handle large-scale systems that are deployable to real-world scenarios across fundamental scientific disciplines (materials, physics, biology, healthcare). Secondly, I intend to integrate external knowledge and signals into GraphODE for more precise and robust/reliable reasoning when faced with limited observational data, spanning from injecting physical constraints, multi-modal learning (e.g. visions, languages) to knowledge graphs. Finally, I plan to explore broader applications of relational reasoning and modeling in real-world contexts. By pursuing these research directions, I aim to push the boundaries of graph-based machine learning in scientific discovery, facilitate more accurate predictions and reasoning, and unlock new possibilities for understanding and leveraging complex dynamical system data in the real world.

Bio:

Zijie Huang is a final-year CS Ph.D. student at UCLA, co-advised by Prof.Yizhou Sun and Prof. Wei Wang. Her research interest lies in graph machine learning in general, with a special focus on reasoning over scientific dynamical systems (GraphML + Science) and knowledge graph modeling. Her research is generously supported by Amazon Ph.D. Fellowship. Her work has been published in top-tier conferences such as Neurips, ICML, KDD, ACL, WWW, and has won the best paper award at the DLDE (The Symbiosis of Deep Learning and Differential Equations) workshop at Neurips 2023. Previously, she interned at Amazon Science and Netflix Research.