Graph Knowledge Transfer: Methods, Practices, and the Knowledge Bridge Learning Framework

Graph Knowledge Transfer Overview

This article shares attempts and viewpoints on knowledge transfer over complex graphs, organized into four main parts: an introduction to graph knowledge transfer, preliminary practical experiments, a universal method called Knowledge Bridge Learning (KBL), and a Q&A session.

1. Introduction to Graph Knowledge Transfer

Graphs, composed of nodes and edges, can represent many real‑world data such as citation networks, social networks, traffic networks, protein interaction networks, and molecular graphs. The Data Hungry problem arises when the amount of usable data is limited or its quality is low, which hampers deep learning methods that require large, high‑quality datasets.

To alleviate this, the concept of open domain data is introduced: abundant data that is not directly usable due to distribution differences. Leveraging such data to supplement target‑domain knowledge can mitigate the Data Hungry issue.

2. Distribution Shift in Graph Data

Open domain data often exhibits significant distribution shift compared to target‑domain data. In graphs, this shift appears at three levels:

Node‑level: differences in node features and types.

Edge‑level: variations in edge features and types.

Graph‑level: overall structural differences between graphs.

Neighbor distribution is also crucial because graph nodes are interconnected. Simple statistics like the mean are insufficient; multi‑order moments are used to better characterize neighbor feature distributions.

3. Traditional Transfer Learning

Traditional domain adaptation separates data into source and target domains and includes scenarios such as zero‑shot, few‑shot, and weak‑shot learning. However, directly transferring classification boundaries often leads to negative transfer when source‑domain knowledge is not fully relevant to the target domain.

4. Practical Graph Knowledge Transfer Experiments

Two real‑world scenarios are explored:

Twitter political election network: political figures (large‑V nodes) have rich features, while ordinary users (small‑V nodes) have sparse information. Knowledge is transferred from large‑V to small‑V nodes to predict user political inclination.

Financial network: listed companies (large‑V) provide comprehensive financial reports, whereas non‑listed companies (small‑V) lack such data. Transferring knowledge improves risk prediction for non‑listed firms.

These scenarios are abstracted into the VS‑Graph (Vocal‑Silent Graph) where vocal nodes are well‑described (large‑V) and silent nodes are sparse (small‑V). The goal is to transfer knowledge across these two node types.

5. Knowledge Transfer on Graph Neural Networks (KTGNN)

The KTGNN pipeline consists of three parts: (1) completing missing attributes of small‑V nodes, (2) message passing on VS‑Graph that respects domain differences, and (3) a domain‑aware classifier for final predictions.

6. Knowledge Bridge Learning (KBL)

KBL extends graph knowledge transfer to three scenarios:

Non‑graph data (e.g., images, vectors, text) where samples are independent.

Cross‑network transfer where two separate graphs belong to different domains.

Multiple domains within a single graph (the VS‑Graph case).

The framework learns a knowledge bridge —edges linking useful source samples to target samples—by constructing a bridged‑graph based on learned similarity scores.

7. Bridged‑GNN Model

The model follows three stages:

Adaptive Knowledge Retrieval (AKR): learns similarity between any pair of samples (intra‑ and inter‑domain) using a decoder and a domain‑divergence learner with adversarial alignment to obtain a delta that aligns target features to the source space.

Bridged‑graph Construction: high‑similarity pairs are connected, while low‑similar edges are pruned, yielding a graph that defines the scope of knowledge transfer.

Graph Neural Network Module: any message‑passing GNN (e.g., KTGNN) operates on the bridged‑graph to propagate knowledge.

Key modules are the knowledge retrieval component, the similarity learning module (including adversarial domain alignment), and the GNN propagation module.

8. Experimental Results

Experiments on ten datasets across the three KBL scenarios demonstrate consistent improvements over baseline transfer learning methods and standard GNNs, with gains of up to 5% absolute accuracy in some cases.

9. Q&A

Q1: Does constructing the bridged‑graph require computing similarity for all sample pairs, and how is efficiency handled?

A1: Similarity is learned using a mini‑batch scalable algorithm rather than full‑batch pairwise computation, and retrieval is accelerated with engineering optimizations.

Q2: What are future directions for graph transfer learning?

A2: Future work includes unsupervised domain adaptation, domain generalization, inductive learning, dynamic graphs, and more complex graph‑based transfer scenarios.

Thank you for your attention.