Bridging Domain Adaptation and Graph Neural Networks: A Tensor-Based Framework for Effective Label Propagation

TL;DR

This work tackles the challenge of transferring graph classifications across domains by proposing LP-TGNN, a tensor-based framework that fuses holistic graph representations with domain adaptation. It combines a Graph Convolutional Branch and a Topological Learning Branch through a Tensor Transformation Layer, enabling effective integration of structural and topological information. Label propagation with pseudo-labels and a FixMatch-style consistency loss regularizes cross-domain predictions, improving transferability without imposing strict domain-invariant constraints. Empirical results across multiple graph benchmarks show LP-TGNN and its ablations outperform baselines, underscoring the value of multi-modal graph representations and topology-aware propagation for domain-adaptive graph classification.

Abstract

Graph Neural Networks (GNNs) have recently become the predominant tools for studying graph data. Despite state-of-the-art performance on graph classification tasks, GNNs are overwhelmingly trained in a single domain under supervision, thus necessitating a prohibitively high demand for labels and resulting in poorly transferable representations. To address this challenge, we propose the Label-Propagation Tensor Graph Neural Network (LP-TGNN) framework to bridge the gap between graph data and traditional domain adaptation methods. It extracts graph topological information holistically with a tensor architecture and then reduces domain discrepancy through label propagation. It is readily compatible with general GNNs and domain adaptation techniques with minimal adjustment through pseudo-labeling. Experiments on various real-world benchmarks show that our LP-TGNN outperforms baselines by a notable margin. We also validate and analyze each component of the proposed framework in the ablation study.

Figures (3)

• Figure 1: A visualization of domain adaptation on graph classification. The red line is the decision boundary of an Enzyme versus Non-Eyzyme classifier trained on $S_E$ and $S_{NE}$. The labels can transfer from the source domain to the target domain via methods such as label propagation pmlr-v139-cai21b. For effective label propagation between domains, the intra-class distance is supposed to be smaller than the inter-class distance NEURIPS2022_ac112e8f. For instance, $T_E$ should be closer to $S_E$ than to $S_{NE}$.
• Figure 2: An illustration of representations produced by the TTG-NN and the state-of-the-art GIN xu2018powerful on the MUTAG dataset Debnath1991-wr. While both graph encoders separate the three clusters effectively, TTG-NN well preserves the in-group distance with green points ($T_0$) and blue points ($S_0$) and pushes away data from a different group: red points ($S_1$). Embeddings plotted using t-SNE JMLR:v9:vandermaaten08a.
• Figure 3: The architecture of the proposed LP-TGNN. The graph mini-batches from both domains are input into both branches. The representations of target samples from each branch are pseudo-labeled by an MLP classifier and each branch is separately regularized by label propagation. Also, the representations of source samples from both branches are concatenated and labeled by the MLP classifier. The model is optimized by the supervised loss and label propagation jointly.

Theorems & Definitions (1)

• Definition 1: Persistence Image