Graph Knowledge Transfer and the Knowledge Bridge Learning Framework

The talk begins with an introduction to knowledge transfer on graphs, defining the problem and highlighting the data‑hungry issue where real‑world graph data often suffer from limited quantity or low quality, making conventional deep learning difficult to apply.

It then describes the concept of open‑domain data and how distribution shifts between source and target domains affect graph nodes, edges, and neighbor structures, motivating the need for specialized graph transfer methods.

Four main parts are covered: (1) an overview of graph knowledge transfer, (2) preliminary experiments on real‑world graph datasets, (3) the Knowledge Bridge Learning (KBL) framework that builds a bridged‑graph to restrict transfer scope, and (4) a Q&A session.

KBL is defined as learning a knowledge bridge—edges linking source and target samples—by first learning similarity across domains, constructing a bridged‑graph based on high‑similarity pairs, and then applying a graph neural network (GNN) to propagate knowledge. The Bridged‑GNN model consists of three modules: an adaptive knowledge retrieval (AKR) module, a similarity learning module with domain‑divergence alignment, and a GNN message‑passing module.

Extensive experiments on non‑graph data, cross‑network scenarios, and single‑graph multi‑domain settings across ten datasets demonstrate that Bridged‑GNN consistently outperforms baseline transfer learning and GNN methods, achieving up to 5% absolute improvement.

The discussion concludes with future work directions, including unsupervised domain adaptation, domain generalization, inductive learning, dynamic graphs, and more advanced GNN architectures for graph transfer learning.