Mitigating Majority Bias in Graph Neural Networks

This direction is merged into another direction: class imbalance in graph learning. Its last version stays readable and citable here.

Version 1 of 151 papers52 conceptsfrozen 2026-10-04

In many real-world networks, the most critical nodes are also the rarest. In financial systems, fraudulent transactions are a tiny fraction of total activity; in biological networks, specific disease-causing proteins are vastly outnumbered by healthy ones. When Graph Neural Networks (GNNs) are trained on these imbalanced datasets, they often fall into a "majority bias" trap. They achieve high overall accuracy by effectively ignoring the minority classes, which are frequently the primary targets of the analysis.

Recent research reveals that this is not just a data-labeling problem, but a structural one inherent to how GNNs process information.

The Blurring Effect of Message Passing

The core mechanism of a GNN is message passing: a node updates its representation by aggregating features from its neighbors. In an imbalanced graph, a minority-class node is statistically likely to be surrounded by majority-class neighbors.

Liang Yan and colleagues identify this as Geometric Imbalance (GI). They argue that message passing "blurs" minority nodes into the majority cluster in the embedding space. As the network aggregates information, the unique signal of a rare node is overwhelmed by the dominant features of its neighbors. Jun Yin describes this as the minority signal being "drowned out," leading to a loss of distinct representations for critical entities like fraudulent transactions.

Beyond Simple Resampling

In traditional machine learning, we might fix imbalance by simply duplicating minority examples (oversampling). In graphs, this is harder because nodes do not exist in isolation; their identity is defined by their connections.

To address this, researchers have moved toward topology-aware augmentation. Lirong Wu introduced GraphMixup, which generates synthetic minority nodes in a "disentangled semantic space" and uses reinforcement learning to determine how many new connections to forge. Similarly, Kai Huang proposed SG-CGAN, which uses Generative Adversarial Networks to create entire synthetic subgraphs for minority classes, ensuring that the "hard-to-classify" structural patterns are preserved.

However, synthetic data can introduce noise. Fanlong Zeng developed GraphIFE, which uses invariant learning to ensure the model focuses on stable, predictive features rather than the potentially low-quality artifacts of synthetic nodes.

The Topological Paradigm

A newer school of thought suggests that we can fix imbalance without rebalancing the data at all. Zhining Liu proposes a topological paradigm called BAT. Instead of adding nodes, BAT identifies and rectifies "Ambivalent" and "Distant" message-passing—instances where the graph structure itself misleads the model.

Another intuitive fix involves the "length" of node vectors. Shenzhi Yang discovered that minority or out-of-distribution nodes often have much smaller representation vectors (norms) than majority nodes. Their framework, NodeReg, enforces norm consistency, essentially forcing the GNN to treat all node vectors as equally important regardless of their class frequency. This prevents the model from "shrinking" minority representations into insignificance.

Adaptive Ensembles and Subclassing

If the majority class is too large and diverse, it can act as a monolithic wall that the GNN cannot see past. Wei Ju addresses this in C3GNN by adaptively clustering majority classes into smaller, semantically coherent subclasses. By breaking the "Goliath" majority into smaller groups, the model can learn fine-grained structural patterns that were previously obscured.

For complex graph classification tasks, GraphDIVE employs a Mixture of Diverse Experts. By viewing the graph through multiple "topological lenses," different parts of the model can specialize in identifying rare structural motifs, such as the specific molecular arrangements of active drugs in a large chemical database.

Learning through a Curriculum

Finally, the order of learning matters. Lijie Wen and Fuli Feng developed GNN-CL, which uses curriculum learning. The model starts by learning easy, well-represented nodes and gradually introduces harder, rarer examples. This prevents the model from prematurely converging on a majority-biased solution, allowing it to build a foundational understanding of the graph before tackling the nuances of the minority classes.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_mitigating_majority_bias_in_graph_neural_006319ce,
  title = {Mitigating Majority Bias in Graph Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-majority-bias-in-graph-neural-networks/rdr_006319ce4cdd4634bd0393cd93ac58e5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_mitigating_majority_bias_in_graph_neural_006319ce_v1,
  title = {Mitigating Majority Bias in Graph Neural Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-majority-bias-in-graph-neural-networks/rdr_006319ce4cdd4634bd0393cd93ac58e5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}