Graph Transfer Learning and VS-Graph: Knowledge Transferable Graph Neural Networks

Graph transfer learning has become widely studied in computer vision and natural language processing, and researchers are now extending it to non‑Euclidean graph data. This article provides an overview of graph transfer learning, discusses challenges caused by distribution shifts across domains, and outlines typical graph representation learning tasks at node, edge, and graph levels.

The new VS-Graph scenario is introduced, where a graph contains two types of nodes: vocal (large, well‑observed) nodes and silent (small, partially observed) nodes. Real‑world examples include political figures versus ordinary users in social networks and listed versus non‑listed companies in financial networks. The goal is to predict labels for silent nodes by transferring knowledge from vocal nodes.

To address this, the Knowledge Transferable Graph Neural Network (KTGNN) is proposed. It consists of three stages:

Domain‑Adaptive Feature Completion (DAFC): Missing attributes of silent nodes are filled using features from neighboring vocal nodes, with domain‑difference correction and attention‑based neighbor importance.

Domain‑Adaptive Message Passing (DAMP): Graph neural network message passing is performed both within domains and across domains, incorporating domain‑difference correction and neighbor importance at each step.

Domain‑Transferable Classifier (DTC): Separate classifiers are trained for vocal and silent nodes; parameters from the vocal classifier are transformed to initialize a target‑domain classifier for silent nodes, with KL‑divergence regularization to preserve domain‑specific knowledge.

Extensive experiments on Twitter social network data and a financial company network demonstrate that KTGNN achieves significant improvements in F1‑score and AUC over baseline methods, even when a large proportion of cross‑domain edges are removed. Ablation studies confirm the contribution of each module.

The paper ("Predicting the Silent Majority on Graphs: Knowledge Transferable Graph Neural Network") is available at https://dl.acm.org/doi/abs/10.1145/3543507.3583287 , and the code is released alongside the publication.

Future research directions include inductive learning for unseen nodes, edge‑level transfer learning, and dynamic VS‑Graphs that evolve over time, requiring models to handle temporal distribution shifts.