Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Transferable Graph Condensation from the Causal Perspective

Huaming Du, Yijie Huang, Su Yao, Yiying Wang, Yueyang Zhou, Jingwen Yang, Jinshi Zhang, Han Ji, Yu Zhao, Guisong Liu, Hegui Zhang, Carl Yang, Gang Kou

TL;DR

This paper addresses the limited transferability of graph dataset condensation by introducing TGCC, a framework that learns causal-invariant features from graph structure and injects them into condensed graphs via spectral-domain contrastive learning. It combines a causal invariant feature extraction module, a graph contrastive condensation objective, and a spectral-domain enhanced contrastive loss to preserve universal causal patterns while maintaining structural and feature information. Experiments across six datasets, including the new FinReport dataset, show state-of-the-art performance in single-task settings and substantial improvements in cross-task and cross-domain transfer, with notable gains such as up to 13.41% on cross-task scenarios. The work implies practical benefits for training efficiency and model generalization in real-world graph learning tasks, and it provides a new resource for the community through FinReport.

Abstract

The increasing scale of graph datasets has significantly improved the performance of graph representation learning methods, but it has also introduced substantial training challenges. Graph dataset condensation techniques have emerged to compress large datasets into smaller yet information-rich datasets, while maintaining similar test performance. However, these methods strictly require downstream applications to match the original dataset and task, which often fails in cross-task and cross-domain scenarios. To address these challenges, we propose a novel causal-invariance-based and transferable graph dataset condensation method, named \textbf{TGCC}, providing effective and transferable condensed datasets. Specifically, to preserve domain-invariant knowledge, we first extract domain causal-invariant features from the spatial domain of the graph using causal interventions. Then, to fully capture the structural and feature information of the original graph, we perform enhanced condensation operations. Finally, through spectral-domain enhanced contrastive learning, we inject the causal-invariant features into the condensed graph, ensuring that the compressed graph retains the causal information of the original graph. Experimental results on five public datasets and our novel \textbf{FinReport} dataset demonstrate that TGCC achieves up to a 13.41\% improvement in cross-task and cross-domain complex scenarios compared to existing methods, and achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario.

Transferable Graph Condensation from the Causal Perspective

TL;DR

This paper addresses the limited transferability of graph dataset condensation by introducing TGCC, a framework that learns causal-invariant features from graph structure and injects them into condensed graphs via spectral-domain contrastive learning. It combines a causal invariant feature extraction module, a graph contrastive condensation objective, and a spectral-domain enhanced contrastive loss to preserve universal causal patterns while maintaining structural and feature information. Experiments across six datasets, including the new FinReport dataset, show state-of-the-art performance in single-task settings and substantial improvements in cross-task and cross-domain transfer, with notable gains such as up to 13.41% on cross-task scenarios. The work implies practical benefits for training efficiency and model generalization in real-world graph learning tasks, and it provides a new resource for the community through FinReport.

Abstract

The increasing scale of graph datasets has significantly improved the performance of graph representation learning methods, but it has also introduced substantial training challenges. Graph dataset condensation techniques have emerged to compress large datasets into smaller yet information-rich datasets, while maintaining similar test performance. However, these methods strictly require downstream applications to match the original dataset and task, which often fails in cross-task and cross-domain scenarios. To address these challenges, we propose a novel causal-invariance-based and transferable graph dataset condensation method, named \textbf{TGCC}, providing effective and transferable condensed datasets. Specifically, to preserve domain-invariant knowledge, we first extract domain causal-invariant features from the spatial domain of the graph using causal interventions. Then, to fully capture the structural and feature information of the original graph, we perform enhanced condensation operations. Finally, through spectral-domain enhanced contrastive learning, we inject the causal-invariant features into the condensed graph, ensuring that the compressed graph retains the causal information of the original graph. Experimental results on five public datasets and our novel \textbf{FinReport} dataset demonstrate that TGCC achieves up to a 13.41\% improvement in cross-task and cross-domain complex scenarios compared to existing methods, and achieves state-of-the-art performance on 5 out of 6 datasets in the single dataset and task scenario.
Paper Structure (21 sections, 1 theorem, 15 equations, 4 figures, 6 tables)

This paper contains 21 sections, 1 theorem, 15 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Graph Dataset Condensation
  4. Causal Inference in Graph Neural Networks
  5. The Proposed Model: TGCC
  6. Overall Framework
  7. Causal Invariant Feature Extraction
  8. Invariance Objective
  9. Independence Objective
  10. Theoretical analysis
  11. Graph Contrastive Condensation
  12. Spectral-domain Enhanced Contrastive Learning
  13. Optimization Objective
  14. Experiments
  15. Experimental Setup
  16. ...and 6 more sections

Key Result

Theorem 1

(Causal Invariance) Given adjacency matrix $A$ and the generated augmentation $V$ , the amplitudes of $i$-th frequency of $A$ and $V$ are ${\lambda }_{i}$ and ${\gamma }_{i}$, respectively. With the optimization of $\mathcal{L}_{causal}$, the following upper bound is established: where ${\theta }_{i}$ is an adaptive weight of the $i$-th term.

Figures (4)

  • Figure 1: The pipeline of existing graph condensation methods and our TGCC method. The main differences between TGCC and existing methods lie in the extraction of causal-invariant features from the graph structure based on causal intervention to achieve transferable graph condensation.
  • Figure 2: An illustrative diagram of the proposed TGCC framework.
  • Figure 3: The generation of negative sample.
  • Figure 4: The accuracy and condensation time of GC methods on Ogbn-Arxiv ($r$ = 0.5%) and FinReport ($r$ = 2.02%).

Theorems & Definitions (1)

  • Theorem 1