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.
Introduction
Graph Neural Networks (GNNs) are currently the gold standard for graph classification. However, their effectiveness is often bottlenecked by a significant "hunger" for labeled data. In domains such as drug discovery or genomics, obtaining labels for a single graph—representing a complex molecule or protein—can require expensive laboratory tests or specialized expert intervention.
Domain Adaptation (DA) offers a potential solution by training models on a labeled "source" domain to perform well on an unlabeled "target" domain. While DA is well-established in computer vision, its application to graph data faces two primary hurdles:
- Holistic Representation: Standard GNNs rely on local message-passing, which can overlook global structural patterns critical for classification.
- Domain Discrepancy: Measuring and aligning the "distance" between distinct graph domains is challenging without losing class-specific discriminative features.
Methodology: The LP-TGNN Framework
To address these challenges, we introduce the Label-Propagation Tensor Graph Neural Network (LP-TGNN). This framework integrates a dual-branch tensor encoder with a label-propagation alignment strategy.
Dual-Branch Tensor Encoder
The model employs two parallel branches to capture a comprehensive representation of graph data:
- Graph Convolutional Branch: Utilizes standard GCN layers to capture local node interactions and neighborhood features.
- Topological Learning Branch: Leverages Persistent Homology (PH) to extract "global shape" features. PH tracks topological features (e.g., holes, connected components) across scales, representing them as Persistent Images (PIs).
These multi-modal features are fused using a Tensor Transformation Layer (TTL). By employing low-rank tensor decomposition (such as Tucker or CP), the TTL efficiently handles high-dimensional data while preserving complex multi-dimensional relationships.
Label Propagation for Domain Alignment
Rather than relying solely on adversarial training, which can inadvertently blur class-specific details, LP-TGNN utilizes Label Propagation. The process involves:
- Generating "pseudo-labels" for target graphs using an MLP classifier trained on source data.
- Regularizing the model to ensure that target graphs share similar representations with source graphs of the same (pseudo) label.
Experimental Evaluation
The framework was evaluated on several real-world benchmarks from the TUDataset, specifically tailored for domain adaptation tasks in bioinformatics (PROTEINS, DD) and chemoinformatics (MUTAG, Mutagenicity, BZR, COX2). To test robustness against synthetic shifts, the authors also created domain discrepancies in the Mutagenicity dataset based on edge density quartiles.
Key Results
LP-TGNN consistently outperformed both standard GNNs and state-of-the-art domain adaptation methods. On average across six domain-shift tasks, LP-TGNN achieved an accuracy of 71.8%, surpassing specialized graph DA methods like CoCo (70.5%) and significantly outperforming vanilla GNNs such as GIN (62.5%).
Discussion and Conclusion
The success of LP-TGNN highlights that global topology serves as a powerful, transferable bridge between domains. While local features may vary, the underlying "global shape" often remains consistent. Furthermore, visualization via t-SNE confirms that the TGNN encoder preserves tighter, more distinct clusters for source and target data compared to standard GNNs.
Future work will focus on scaling the framework to larger datasets and extending the methodology to multi-source or multi-target domain adaptation scenarios. While computationally efficient due to low-rank decomposition, the calculation of Persistent Homology remains a target for further optimization in extremely large-scale graph applications.
Summary updated