Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Propagation for Unsupervised Graph Domain Adaptation

Meihan Liu, Zeyu Fang, Zhen Zhang, Ming Gu, Sheng Zhou, Xin Wang, Jiajun Bu

TL;DR

This work revisits UGDA and reveals that the propagation component of Graph Neural Networks drives cross-domain generalization more than the transformation steps. It proposes A2GNN, an asymmetric architecture that applies a single transformation on the source graph and multiple propagation layers on the target graph, sharing weights to tighten the target risk bound. The authors derive a graph-domain adaptation bound and show that reducing the Lipschitz constant of the feature extractor via propagation-focused design yields a tighter bound, supported by comprehensive experiments on real-world datasets. Empirically, A2GNN outperforms state-of-the-art baselines on various UGDA tasks, and the theoretical results provide guidance for choosing propagation depth and scheme to improve domain adaptation performance.

Abstract

Unsupervised Graph Domain Adaptation (UGDA) aims to transfer knowledge from a labelled source graph to an unlabelled target graph in order to address the distribution shifts between graph domains. Previous works have primarily focused on aligning data from the source and target graph in the representation space learned by graph neural networks (GNNs). However, the inherent generalization capability of GNNs has been largely overlooked. Motivated by our empirical analysis, we reevaluate the role of GNNs in graph domain adaptation and uncover the pivotal role of the propagation process in GNNs for adapting to different graph domains. We provide a comprehensive theoretical analysis of UGDA and derive a generalization bound for multi-layer GNNs. By formulating GNN Lipschitz for k-layer GNNs, we show that the target risk bound can be tighter by removing propagation layers in source graph and stacking multiple propagation layers in target graph. Based on the empirical and theoretical analysis mentioned above, we propose a simple yet effective approach called A2GNN for graph domain adaptation. Through extensive experiments on real-world datasets, we demonstrate the effectiveness of our proposed A2GNN framework.

Rethinking Propagation for Unsupervised Graph Domain Adaptation

TL;DR

This work revisits UGDA and reveals that the propagation component of Graph Neural Networks drives cross-domain generalization more than the transformation steps. It proposes A2GNN, an asymmetric architecture that applies a single transformation on the source graph and multiple propagation layers on the target graph, sharing weights to tighten the target risk bound. The authors derive a graph-domain adaptation bound and show that reducing the Lipschitz constant of the feature extractor via propagation-focused design yields a tighter bound, supported by comprehensive experiments on real-world datasets. Empirically, A2GNN outperforms state-of-the-art baselines on various UGDA tasks, and the theoretical results provide guidance for choosing propagation depth and scheme to improve domain adaptation performance.

Abstract

Unsupervised Graph Domain Adaptation (UGDA) aims to transfer knowledge from a labelled source graph to an unlabelled target graph in order to address the distribution shifts between graph domains. Previous works have primarily focused on aligning data from the source and target graph in the representation space learned by graph neural networks (GNNs). However, the inherent generalization capability of GNNs has been largely overlooked. Motivated by our empirical analysis, we reevaluate the role of GNNs in graph domain adaptation and uncover the pivotal role of the propagation process in GNNs for adapting to different graph domains. We provide a comprehensive theoretical analysis of UGDA and derive a generalization bound for multi-layer GNNs. By formulating GNN Lipschitz for k-layer GNNs, we show that the target risk bound can be tighter by removing propagation layers in source graph and stacking multiple propagation layers in target graph. Based on the empirical and theoretical analysis mentioned above, we propose a simple yet effective approach called A2GNN for graph domain adaptation. Through extensive experiments on real-world datasets, we demonstrate the effectiveness of our proposed A2GNN framework.
Paper Structure (43 sections, 21 equations, 4 figures, 13 tables)

This paper contains 43 sections, 21 equations, 4 figures, 13 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Preliminaries
  4. Notations.
  5. Problem Definition.
  6. Rethinking Propagation for UGDA
  7. Empirical Analysis
  8. Observation 1: Propagation operation reflects the generalization capability of GNNs.
  9. Observation 2: Conducting propagation solely on target graph is of primary importance.
  10. The Proposed Framework
  11. Theoretical Analysis
  12. Theorem 1.
  13. Lemma 1.
  14. Lemma 2.
  15. Lemma 3.
  16. ...and 28 more sections

Figures (4)

  • Figure 1: The influence of different operations in GNNs on task $D \rightarrow A$. $[\cdot]$ indicates the module stacking operation.
  • Figure 2: The framework of existing models and our proposed A2GNN model.
  • Figure 3: (a) shows the sensitivity of the num of propagation layers $k$. (b) shows the sensitivity of trade-off parameter $\alpha$.
  • Figure 4: (a-c) shows the sensitivity of the number of propagation layers $k$. (a) is the Micro-F1 of ${\rm A2GNN\xspace}_{mmd}$. (b-c) is the Macro-F1 and Micro-F1 of ${\rm A2GNN\xspace}_{adv}$ respectively. (d-f) shows the sensitivity of trade-off parameter $\alpha$. (d) is the Micro-F1 of ${\rm A2GNN\xspace}_{mmd}$. (e-f) is the Macro-F1 and Micro-F1 of ${\rm A2GNN\xspace}_{adv}$ respectively.