Abstract
We present Multiple Graph Adversarial Learning (MGAL), a novel framework designed to learn consistent node representations from data characterized by multiple distinct graph structures. By employing an adversarial setup consisting of multiple generators and a single discriminator, MGAL identifies a structure-invariant subspace where node embeddings remain consistent regardless of the specific graph structure from which they originated. This approach significantly enhances performance in semi-supervised learning tasks by effectively capturing both unique structural information and cross-graph correlations.
Introduction
Graph Convolutional Networks (GCNs) have become a standard for relational data. However, most GCN architectures are optimized for a single graph structure. In many real-world scenarios, the same set of nodes is connected by multiple types of relationships—for instance, an image dataset where graphs represent color, texture, and shape similarities. The primary challenge lies in integration: exploiting unique structural information while capturing the underlying correlations between different views. Previous approaches, such as heuristic graph fusion or simple result concatenation, often fail to find a truly consistent representation that captures the semantic essence of nodes across all views.
Methodology
The MGAL framework addresses these challenges through three integrated components:
- The Generator (GCNs): The model utilizes GCN layers with shared parameters across all input graphs. For each graph $A^{(v)}$, the generator produces a latent representation $Z^{(v)}$. Parameter sharing encourages the initial propagation of knowledge across different structures.
- The Discriminator: A Multi-Layer Perceptron (MLP) acts as a discriminator. Its objective is to identify which of the $m$ graphs a specific node representation originated from.
- Adversarial Learning: The generator and discriminator engage in a minimax game. The generator is trained to "fool" the discriminator by creating node embeddings that are so consistent and structure-invariant that the discriminator cannot distinguish their source graph. This forces the model to find a common subspace representing the nodes' true identities.
- Classification Layer: The final representations from all graphs are concatenated ($Z = [Z^{(1)} || Z^{(2)} ...]$) and passed through a fully connected layer for semi-supervised node classification.
Experimental Results
The authors evaluated MGAL on four diverse datasets: Handwritten Numerals, Caltech101-7, MSRC-v1, and CiteSeer. MGAL consistently outperformed several baselines, including single-graph GCNs and averaged-graph GCNs (GCN-M).
- Accuracy Gains: On the MSRC-v1 dataset with a 10% label ratio, MGAL achieved 88.68% accuracy, significantly higher than the best single-graph GCN (81.21%).
- Robustness: Performance generally improved as the number of integrated graphs increased, demonstrating the model's ability to effectively aggregate information from multiple sources.
- Qualitative Analysis: t-SNE visualizations confirm that MGAL produces more distinct and tighter clusters for different classes compared to standard GCNs, which often suffer from overlapping class boundaries.
Conclusion
MGAL demonstrates that adversarial learning can serve as a powerful regularizer for multi-graph data. By penalizing graph-specific embeddings, the framework forces GCNs to focus on shared semantic information. While the current model assumes fixed node sets across graphs, future work may extend this to partial multi-graph scenarios and dynamic, time-evolving structures.
Summary updated