Standard Graph Neural Networks (GNNs) typically operate under a "closed-set" assumption, where all nodes encountered during testing are assumed to belong to the classes seen during training. However, real-world graphs are dynamic and often contain novel categories. For instance, a citation network model trained on "Physics" and "Biology" might fail when encountering a "Quantum Computing" paper, often misclassifying it into a known category with high confidence. This paper introduces $G^2Pxy$, a generative framework designed for open-set node classification that accurately identifies known classes while rejecting unknown ones.
Methodology
The $G^2Pxy$ framework operates in a strict inductive setting, meaning it has no prior knowledge of unknown classes during training. It addresses the open-set challenge through two primary mechanisms:
- Proxy Node Generation via Manifold Mixup: Instead of using computationally expensive Generative Adversarial Networks (GANs), $G^2Pxy$ employs manifold mixup in the latent space of the GNN. It generates "proxy unknowns" by interpolating between hidden representations at a middle layer of the network.
- Inter-class Proxies: Created by mixing nodes from different known classes that share an edge, simulating unknowns at the boundaries of known categories.
- External Proxies: Generated by mixing "peripheral" nodes (e.g., leaf nodes or unconfident nodes) with "reversed" class centers to represent unknowns at the fringes of the data distribution.
- Open-Set Classifier Learning: The model extends a standard $C$-class classifier to a $(C+1)$-class classifier, where the additional class represents "unknown." Training utilizes a combination of standard Cross-Entropy and a Tailored Complement Entropy (CoE) loss. This loss function calibrates the model by ensuring the "unknown" class is the second-most likely prediction for known nodes, effectively sharpening the decision boundaries between known and novel data.
Experimental Evaluation
The researchers evaluated $G^2Pxy$ on four standard citation network benchmarks: Cora, Citeseer, DBLP, and PubMed. They tested the model's performance in two specific scenarios:
- Near Open-Set: One class within the dataset is held out as the "unknown" during training.
- Far Open-Set: Nodes from entirely different datasets are used as the "unknowns" (e.g., using Cora as known and Citeseer as unknown).
Key Results and Insights
$G^2Pxy$ consistently outperformed state-of-the-art baselines, including PROSER and OpenWGL, across all datasets. In inductive "near open-set" tasks, it achieved an average 3.70% improvement in accuracy and a significant 9.60% improvement in Macro-F1 score over the second-best method.
Several key insights emerged from the study:
- Efficient Generation: The paper demonstrates that manifold mixup provides a lightweight yet effective alternative to heavy generative machinery for synthesizing unknown distributions.
- Boundary Awareness: By targeting boundary and peripheral nodes for proxy generation, the model learns to define the limits of the "known" space much more effectively than global thresholding methods.
- Calibration via Complement Entropy: The CoE loss prevents the overconfidence typical of deep networks by forcing a more nuanced relationship between known and unknown classes.
Conclusion
$G^2Pxy$ represents a significant step forward in making GNNs more robust for real-world applications where unknown data is inevitable. While highly effective, future research could address scaling the method to massive graphs with millions of nodes, exploring non-linear proxy synthesis techniques, and adapting the framework for dynamic graphs where new classes emerge continuously over time.
Summary updated