Graph Neural Networks (GNNs) have achieved remarkable success in modeling relational data, from molecular structures to social networks. However, their performance heavily relies on an abundance of high-quality, manually annotated labels. In specialized domains like drug discovery, neuroscience, or financial fraud detection, acquiring these labels requires expensive laboratory experiments or rare expert knowledge.
When labels are scarce, standard GNNs suffer from severe performance degradation and overfitting. Because GNNs learn by aggregating messages from a node's local neighborhood, a lack of supervision causes the model to memorize specific local patterns rather than learning generalizable structural features.
To overcome this bottleneck, a prominent research direction focuses on a core intuition: graphs contain explicit structural topology that can substitute for missing semantic labels. By leveraging properties like homophily (the tendency of connected nodes to share attributes) and structural invariance, models can learn robust representations even in low-data regimes. This article distills how researchers use self-supervision, topology-aware active learning, and cross-domain transfer to make graph learning data-efficient.
Structural Self-Supervision via Contrastive Learning
When external labels are missing, the graph itself can provide a rich source of internal supervision. Graph Contrastive Learning (GCL) exploits this by creating multiple "views" of the same graph through structural perturbations (such as dropping edges or masking node features) and training the network to recognize that these views originate from the same source.
In Graph Contrastive Learning with Augmentations, Yuning You demonstrated that maximizing the mutual information between different augmented views forces the GNN to capture high-level, invariant structural properties. This prevents the model from overfitting to the specific training instances that happen to be labeled.
Original Graph ----> Augmentation A (Drop Edges) ----\
----> Augmentation B (Mask Features) ---> Maximize Mutual Information
To maximize the utility of self-supervision, Hassani et al. introduced a multi-view approach in Contrastive Multi-View Representation Learning on Graphs that contrasts a graph's local structure (its adjacency matrix) with its global context (computed via graph diffusion).
Why it works
Contrasting local and global views forces the node embeddings to encode both immediate neighborhood relationships and broader topological roles simultaneously. By learning what makes a graph structurally unique before any human labels are introduced, the model constructs a highly regularized representation space. When a handful of downstream labels are finally provided, the model only needs to learn a simple mapping from this robust space to the target classes.
Topology-Driven Active Learning
If budget constraints allow for only a tiny fraction of a graph to be labeled, which nodes should we choose? Traditional active learning strategies select instances based on model uncertainty. On graphs, however, an uncertain node might be an isolated outlier; labeling it provides no useful information about the rest of the network.
To solve this "cold-start" problem, Xu et al. proposed FeatProp in Active Learning for Graph Neural Networks via Node Feature Propagation. Instead of evaluating nodes in isolation, FeatProp propagates and smooths features across the graph topology before applying a clustering algorithm (like K-Medoids) to select nodes for labeling.
Why it works
Feature propagation explicitly aligns the node features with the graph's connectivity matrix. When you cluster these smoothed features, the cluster centers ("medoids") naturally correspond to nodes that are both feature-representative and structurally central. Labeling these central nodes maximizes the reach of the GNN's message-passing mechanism, allowing a single label to effectively inform an entire community of surrounding unlabeled nodes.
Taking this concept further, Zhang et al. framed node selection as a social influence maximization problem in RIM: Reliable Influence-based Active Learning on Graphs. Their framework, RIM, evaluates both the quantity of a node's structural influence over its neighbors and the quality (reliability) of its potential label, ensuring that the selected nodes act as optimal information hubs during training.
Cross-Domain Knowledge Transfer
When a target graph has virtually no labels, an alternative strategy is to transfer knowledge from a label-rich source graph, even if it comes from an entirely different domain (e.g., transferring patterns from a well-studied social network to a newly mapped biological network).
In Graph Few-Shot Learning via Knowledge Transfer, Huaxiu Yao introduced a meta-learning framework that transfers structural knowledge via a prototype-based GNN. By learning a hierarchical gating mechanism on auxiliary graphs, the model learns how to adapt its internal representations rapidly when encountering a new task with only a few examples.
Label-Rich Source Graph ----> Learn Universal Topological Patterns ----\
|---> Fast Adaptation
Label-Scarce Target Graph --> Extract Local Structure -----------------/
However, transferring knowledge across graphs often introduces a "domain gap" because node features and edge densities can vary drastically. To bridge this gap, Wen et al. developed LP-TGNN in Bridging Domain Adaptation and Graph Neural Networks: A Tensor-Based Framework for Effective Label Propagation. Their framework uses a tensor-based architecture to capture holistic graph information alongside a label propagation mechanism designed to minimize domain discrepancies.
Why it works
While the semantic meaning of node features might change between domains, the underlying geometric and topological principles—such as community clustering, degree distributions, and relative node distances—often remain invariant. By using label propagation as a structural regularizer during domain adaptation, the model smooths out feature discrepancies, allowing the structural insights learned from the source graph to generalize robustly to the low-resource target graph.