Abstract
Graph Neural Networks (GNNs) often suffer from performance degradation on imbalanced datasets, where majority classes dominate the learning process. This paper introduces NeuBM (Neutral Bias Mitigation), a novel post-processing calibration technique. NeuBM identifies inherent model bias by evaluating a statistically neutral reference graph and subtracts this bias from the final predictions. Experimental results across eight benchmark datasets demonstrate that NeuBM significantly boosts minority class performance with minimal computational overhead and no requirement for model retraining.
Introduction
In real-world applications like fraud detection or rare disease identification, graph data is frequently imbalanced. This "topology imbalance" allows majority class influence to propagate through edges, skewing decision boundaries and marginalizing minority nodes. Traditional solutions like resampling or loss weighting often disrupt graph structure or require complex hyperparameter tuning. NeuBM addresses this by correcting model bias post-hoc, preserving the integrity of the original learned representations.
Methodology: Neutral Input Calibration
NeuBM operates through a two-step process:
- Neutral Graph Construction: A synthetic "neutral" graph ($G_{neutral}$) is generated to represent the statistical average of the dataset. It is constructed using the mean and covariance of node features and the average edge density, serving as a class-agnostic baseline.
- Logit Calibration: The pre-trained GNN processes $G_{neutral}$ to produce "neutral logits" ($L_{neutral}$), which encapsulate the model's inherent preference for majority classes. The final prediction for a target graph $G$ is then calibrated by subtraction: $$L_{corrected} = f_\theta(G) - f_\theta(G_{neutral})$$
This simple linear offset effectively shifts the decision boundary to a more equitable position.
Experimental Results
The authors evaluated NeuBM on eight datasets, including citation networks (Cora, Citeseer, PubMed) and e-commerce graphs (Amazon Computers, Amazon Photo).
Performance Gains
NeuBM consistently outperformed state-of-the-art methods like TAM and GraphSMOTE. On the Cora dataset, NeuBM improved the F1-score of the smallest minority class by 26.9% compared to the best baseline.
Scalability and Efficiency
Despite the additional forward pass, NeuBM remains highly efficient. Across various datasets, it maintains a superior F1-macro score while adding only marginal computation time (approximately 1.3x).
Conclusion
NeuBM offers a practical and "plug-and-play" solution for mitigating bias in GNNs. By treating imbalance as a calibration challenge rather than a data-level problem, it achieves significant fairness improvements without the need for expensive retraining. Future work includes extending this approach to dynamic graphs and extremely large-scale networks.
Summary updated