Semi-supervised Graph Neural Network
Semi-supervised Graph Neural Network (GNN with Label Propagation) · Also known as: Semi-supervised GNN, GNN semi-supervised learning, graph-based semi-supervised classification, semi-supervised node classification
A semi-supervised graph neural network trains a GNN on a graph where only a small fraction of nodes carry labels, using neighborhood message-passing to spread information from labeled nodes to unlabeled ones. The approach, popularised by Kipf and Welling's 2017 Graph Convolutional Network, achieves strong node-classification accuracy even when labeled examples are scarce.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use a semi-supervised GNN when your data are naturally graph-structured — citation networks, social networks, knowledge graphs, molecular graphs, user-item interaction graphs — and labeled examples are expensive or sparse (typically under 5% of nodes). It outperforms purely supervised GNNs when labels are scarce and outperforms classical label-propagation when node features are informative. Do not use it when the graph is very small (fewer than a few hundred nodes), the homophily assumption is weak (connected nodes do not tend to share the same class), or when no reliable graph structure exists and a constructed k-NN graph would be noisy. Tabular data without a natural graph structure is usually better served by standard supervised models.
Strengths & limitations
- Exploits both graph topology and node features, combining two complementary information sources.
- Achieves competitive accuracy with very few labeled examples — sometimes only 20 labels per class.
- Scales to millions of nodes with mini-batch variants such as GraphSAGE and GraphSAINT.
- Naturally handles node-feature heterogeneity and varying node degrees.
- Representations learned are reusable for downstream tasks beyond node classification, including link prediction and graph classification.
- Transductive GCN requires the full graph at training time and cannot trivially predict for completely isolated new nodes.
- Performance degrades on heterophilic graphs where connected nodes belong to different classes.
- Deeper GNNs suffer from over-smoothing: node representations become indistinguishable after many message-passing layers.
- Constructing a meaningful graph from non-graph data introduces an additional modeling choice that strongly influences results.
Frequently asked
How many labeled nodes are typically needed?
Kipf and Welling demonstrated strong results with as few as 20 labeled nodes per class on standard citation benchmarks, though performance improves steadily as more labels are added. The minimum depends on graph size, feature informativeness, and class separability.
How does this differ from a fully supervised GNN?
A fully supervised GNN computes loss only over labeled nodes but is assumed to have a large labeled set. Semi-supervised GNNs are specifically designed and evaluated for the regime where labeled nodes are a tiny fraction of the graph, and they explicitly leverage unlabeled neighbors' structural context.
What if my graph is heterophilic?
Standard GCN-style aggregation assumes connected nodes share the same class. On heterophilic graphs, aggregating neighbors can hurt performance. Architectures like GPRGNN, H2GCN, or signed-graph GNNs are better suited for heterophilic settings.
Can I apply this to data that is not naturally a graph?
Yes, by constructing a k-nearest-neighbor graph on feature similarity. However, the quality of the learned representations depends heavily on whether the resulting graph reflects meaningful structure. Noisy constructed graphs can hurt rather than help.
How do I prevent over-smoothing?
Use shallow architectures (2–3 layers), add residual connections, apply DropEdge or PairNorm regularization, or use jumping knowledge networks that aggregate representations from all layers rather than only the final one.
Sources
- Kipf, T. N., & Welling, M. (2017). Semi-Supervised Classification with Graph Convolutional Networks. International Conference on Learning Representations (ICLR 2017). link ↗
- Zhou, D., Bousquet, O., Lal, T. N., Weston, J., & Scholkopf, B. (2004). Learning with Local and Global Consistency. Advances in Neural Information Processing Systems (NeurIPS 2004), 17. link ↗
How to cite this page
ScholarGate. (2026, June 3). Semi-supervised Graph Neural Network (GNN with Label Propagation). ScholarGate. https://scholargate.app/en/deep-learning/semi-supervised-graph-neural-network
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- Graph Convolutional NetworkDeep learning↔ compare
- Graph Neural Network (Network Analysis)Network analysis↔ compare
- Label PropagationMachine learning↔ compare
- Semi-supervised LearningMachine learning↔ compare