Many real-world problems are inherently graph-structured, such as social networks, biological networks, the World Wide Web, molecules, circuits, brains, road networks, and knowledge graphs. Many machine learning algorithms are also defined on graphs, such as neural networks and graphical models. In this field, we develop algorithms and theories for learning over graphs, and apply them to problems like link prediction, graph classification, graph structure optimization, and knowledge graph reasoning. We are also interested in practical applications of graph neural networks, including brain modeling, drug discovery, circuit design, and healthcare applications.

Compared to machines, humans possess extreme flexibility in handling unseen tasks in a few-shot/zero-shot way, much of which is attributed to human system-II intelligence for complex logical reasoning, task planning, causal reasoning, and inductive generalization. Large language models (LLMs) have shown unprecedented improvement in such abilities, but still face challenges in top human-level tasks, such as scientific innovation, software engineering, long-form writing, and autonomous agents. In this field, we aim to: a) Design next-generation LLM architectures with long-term memory, human-like learning mechanisms, fast training/inference, and superior long-context abilities. b) Improve LLMs' reasoning ability to match or surpass humans in the most complex tasks. c) Explore LLMs' integration with various modalities, such as graphs, code, relational databases (RDB), images, and videos.