Graph Learning under Distribution Shift: From Continual Learning to Zero-shot Reasoning

The School of EECS is hosting the following HDR Progress Review 3 Seminar:

Graph Learning under Distribution Shift: From Continual Learning to Zero-shot Reasoning

Speaker: Yilun Liu

Abstract: Prediction on attributed graphs underpins citation analysis, recommendation, information retrieval, and knowledge discovery. Unlike independent data instances, graph nodes are interconnected. A node's label is predicted from the attributes of both the node and the surrounding nodes. Existing methods exploit this dependency by jointly modelling node attributes and graph topology.

However, most methods assume that training and test data share the same distribution, label space, and domain semantics. In practice, this assumption rarely holds: graphs evolve, new classes emerge, and graph data appears in previously unfamiliar domains. This seminar​covers graph learning under distribution shift through four directions: learnable memory for continual graph learning, label-efficient learnable memory, language-based zero-shot graph reasoning with LLMs for unfamiliar domains, and dense reward supervision for graph reasoning traces. Together, these approaches aim to retain knowledge of earlier classes, adapt efficiently with limited labels, generalise to unfamiliar domains, and strengthen supervision over reasoning traces.

Bio: Yilun Liu is a PhD candidate at the School of Electrical Engineering and Computer Science, The University of Queensland. His research focuses on graph learning under distribution shift, with an emphasis on continual graph learning and language-based graph reasoning with LLMs.