In standard machine learning, class imbalance means having fewer examples of one category than another—such as having fewer images of cats than dogs. While the model might struggle to learn the features of a cat, the cat images themselves remain distinct.
In graph representation learning, however, data points are interconnected. When a graph exhibits class imbalance, minority nodes are not just rare; they are structurally overwhelmed. Because graph neural networks (GNNs) rely on message passing to aggregate information from local neighborhoods, a rare node is frequently surrounded by a sea of majority-class neighbors. This structural entanglement creates a unique failure mode where minority classes are literally absorbed by the majority.
The Mechanism of Neighborhood Smearing
Why do standard GNNs fail so dramatically under class imbalance? The root cause lies in the interaction between network topology and the message-passing mechanism.
GNNs update a node's representation by aggregating features from its neighbors. As Yan et al. demonstrate in their work on geometric imbalance, this aggregation process acts as a force that pulls minority-class node representations directly toward majority-class clusters in the embedding space.
This phenomenon, termed "minority-class assimilation" by Wang et al., occurs because underrepresented classes often exhibit lower homophily ratios. As Liang observes, rare nodes are predominantly surrounded by neighbors from different (majority) classes. When a GNN convolves over these neighborhoods, the overwhelming volume of majority-class messages drowns out the distinct signal of the minority node. As a result, the unique features of minority classes are progressively compressed into narrow, squeezed latent subspaces (Li et al.), making them nearly indistinguishable from the majority class.
Standard Message Passing in Imbalanced Graphs:
[Majority Node] \
[Majority Node] --> [Minority Node] == (Message Passing) ==> Representation pulled
[Majority Node] / toward Majority Cluster
Neighbor Memorization and Topology Imbalance
Even when a model attempts to learn a minority class, it faces a structural trap: overfitting to local topology. Because there are so few minority examples, the model encounters only a handful of neighborhood configurations. Park et al. identify this as "neighbor memorization," where the GNN overfits to the specific, limited local structures of minority classes rather than learning generalizable node features.
Crucially, this problem extends beyond simple node counts. Zhao et al. reveal a deeper layer of the problem called "topology-level imbalance." They show that even when class labels are perfectly balanced, significant performance gaps emerge if certain structural motifs or graph shapes within a class are rare. The model prioritizes dominant topological configurations, causing it to fail on infrequent but equally valid sub-structures.
Countering Imbalance via Structural Intervention
To break majority-class dominance, recent research has shifted from simple loss re-weighting to active structural and geometric interventions.
Generative Augmentation of Ego Networks
Traditional oversampling methods like SMOTE interpolate features in isolation. In graphs, however, a newly synthesized node needs edges to participate in message passing. Zhao et al. addressed this with GraphSMOTE, which synthesizes minority-class nodes in the embedding space alongside an edge generator that dynamically wires them into the network.
To combat neighbor memorization, Park et al. took this further with GraphENS, a framework that synthesizes entire ego networks (the target node and its complete neighborhood) rather than isolated nodes. This provides the GNN with diverse structural contexts, preventing it from memorizing fixed local topologies.
Structural and Geometric Filtering
Rather than just adding data, models must actively resist assimilation during training. The UNREAL framework by Yan et al. uses geometric filtering and dual-path pseudo-labeling to explicitly counteract neighborhood-induced smearing, isolating minority signals before they are overwhelmed. Similarly, Wang et al. utilize a relation diffusion strategy to structurally balance the graph topology before synthesis occurs, ensuring that minority representations remain distinct during message passing.
Subclass Decomposition
Another intuitive angle is to break down the monolithic majority class. Standard GNNs exhibit a "greedy" learning behavior, optimizing for the abundant majority data at the expense of sparse minority classes. Ju et al. introduce C3GNN, which adaptively clusters majority classes into smaller, semantically coherent subclasses. By treating the dominant class as a collection of distinct sub-groups, the model preserves structural diversity and prevents a single majority pattern from overpowering the global training signal.