NeuBM: Mitigating Model Bias in Graph Neural Networks through Neutral Input Calibration

Jiawei Gu, Ziyue Qiao, Xiao Luo

Abstract

Graph Neural Networks (GNNs) have shown remarkable performance across various domains, yet they often struggle with model bias, particularly in the presence of class imbalance. This bias can lead to suboptimal performance and unfair predictions, especially for underrepresented classes. We introduce NeuBM (Neutral Bias Mitigation), a novel approach to mitigate model bias in GNNs through neutral input calibration. NeuBM leverages a dynamically updated neutral graph to estimate and correct the inherent biases of the model. By subtracting the logits obtained from the neutral graph from those of the input graph, NeuBM effectively recalibrates the model's predictions, reducing bias across different classes. Our method integrates seamlessly into existing GNN architectures and training procedures, requiring minimal computational overhead. Extensive experiments on multiple benchmark datasets demonstrate that NeuBM significantly improves the balanced accuracy and recall of minority classes, while maintaining strong overall performance. The effectiveness of NeuBM is particularly pronounced in scenarios with severe class imbalance and limited labeled data, where traditional methods often struggle. We provide theoretical insights into how NeuBM achieves bias mitigation, relating it to the concept of representation balancing. Our analysis reveals that NeuBM not only adjusts the final predictions but also influences the learning of balanced feature representations throughout the network.

Introduction

In real-world applications of Graph Neural Networks (GNNs), such as fraud detection, rare disease diagnosis, and social network analysis, data is rarely distributed evenly across classes. This inherent class imbalance poses a significant challenge: GNNs tend to develop a "majority rules" bias. Because nodes in a graph are interconnected, the influence of majority classes propagates through edges during the message-passing process, often drowning out the signals of underrepresented minority classes.

Existing solutions to this problem typically involve complex data resampling—which can inadvertently disrupt the graph's structural integrity—or computationally expensive retraining. NeuBM (Neutral Bias Mitigation) addresses these challenges by introducing a post-processing calibration technique that identifies and subtracts inherent model bias without requiring architectural overhauls or extensive retraining.

Methodology: The Neutral Input Calibration Framework

NeuBM operates on the principle that a model's bias can be quantified by observing its response to a statistically "neutral" input. The framework consists of three primary stages:

1. Neutral Graph Construction

The researchers construct a synthetic "Neutral Graph" ($G_{neutral}$) designed to represent the average state of the dataset. This graph matches the average node count and edge density of the training data. Crucially, its node features are sampled from a multivariate Gaussian distribution based on the global mean and covariance of the entire dataset. This ensures the graph is class-agnostic and serves as a baseline for the model's default behavior.

2. Logit Extraction

The pre-trained GNN processes this neutral graph to produce "neutral logits." These logits represent the model's inherent preference—essentially, what the model predicts when presented with non-descript, average data. If the model predicts a specific class even for this neutral input, it indicates a structural bias toward that class.

3. Calibration via Subtraction

When the model encounters a real input graph, it generates standard logits ($L$). NeuBM applies a correction formula:

$$L_{corrected} = L - L_{neutral}$$

By subtracting the neutral logits, the method offsets the overconfidence typically directed toward majority classes and mitigates the negative shifts that disadvantage minority classes. This "subtraction trick" effectively recalibrates the decision boundary to be more equitable across all classes.

Experimental Evaluation

The effectiveness of NeuBM was evaluated across eight diverse benchmark datasets with imbalance ratios ranging from 2 to 18, including citation networks (Cora, Citeseer, PubMed, Cora-ML, DBLP), e-commerce networks (Amazon Computers, Amazon Photo), and social networks (Twitch PT).

Performance and Scalability

NeuBM consistently outperformed state-of-the-art baselines such as TAM, ReNode, and GraphENS. On the Cora dataset, NeuBM improved the F1-macro score—a critical metric for imbalanced data—from approximately 0.60 to 0.71. The most significant gains were observed in extreme imbalance scenarios; for the smallest minority class in the Cora dataset, NeuBM achieved a 26.9% improvement in F1-score over the best-performing baseline.

Furthermore, the method proved to be architecture-agnostic, showing improvements across GCN, GAT, GraphSAGE, and Graph Transformers. In terms of efficiency, NeuBM adds minimal overhead, requiring only about 1.3x the training time of a standard GCN, making it highly scalable for practical applications.

Theoretical Grounding

The authors provide a theoretical basis for NeuBM by linking it to representation balancing. They demonstrate that the subtraction method reduces the Maximum Mean Discrepancy (MMD) between classes. This suggests that NeuBM does more than just adjust final outputs; it encourages the network to learn more balanced feature representations, effectively neutralizing the structural advantages usually held by majority nodes.

Conclusion and Future Work

NeuBM offers a robust, efficient, and plug-and-play solution for the class imbalance problem in GNNs. By using a neutral baseline to calibrate predictions, it restores visibility to minority classes without the need for destructive data manipulation.

While highly effective, future research may focus on optimizing neutral graph construction for graphs with billions of nodes to avoid memory bottlenecks. Additionally, extending this calibration technique to dynamic or heterogeneous graphs—where node types and structures evolve over time—presents a promising avenue for further mitigating bias in complex, real-world systems.

Summary updated

Abstract

Graph Neural Networks (GNNs) have shown remarkable performance across various domains, yet they often struggle with model bias, particularly in the presence of class imbalance. This bias can lead to suboptimal performance and unfair predictions, especially for underrepresented classes. We introduce NeuBM (Neutral Bias Mitigation), a novel approach to mitigate model bias in GNNs through neutral input calibration. NeuBM leverages a dynamically updated neutral graph to estimate and correct the inherent biases of the model. By subtracting the logits obtained from the neutral graph from those of the input graph, NeuBM effectively recalibrates the model's predictions, reducing bias across different classes. Our method integrates seamlessly into existing GNN architectures and training procedures, requiring minimal computational overhead. Extensive experiments on multiple benchmark datasets demonstrate that NeuBM significantly improves the balanced accuracy and recall of minority classes, while maintaining strong overall performance. The effectiveness of NeuBM is particularly pronounced in scenarios with severe class imbalance and limited labeled data, where traditional methods often struggle. We provide theoretical insights into how NeuBM achieves bias mitigation, relating it to the concept of representation balancing. Our analysis reveals that NeuBM not only adjusts the final predictions but also influences the learning of balanced feature representations throughout the network.
BibTeX
@inproceedings{gu2025neubm,
  author = {Jiawei Gu and Ziyue Qiao and Xiao Luo},
  title = {{NeuBM: Mitigating Model Bias in Graph Neural Networks through Neutral Input Calibration}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.24963/ijcai.2025/315},
}
Figure