Is a GCN the same as every graph neural network?

No. GCNs are one family within the broader graph-neural-network field.

Does one layer use the whole graph for every node?

In the local formulation described here, one layer aggregates from immediate neighbors and the node itself; stacked layers extend the reach.

Are edges learned automatically?

Not necessarily. The standard setup takes a graph as input, although other methods can learn or modify graph structure.

AI Glossary / Graphs and symbolic AI / Intermediate

Graph Convolutional Network

Definition

A Graph Convolutional Network, or GCN, is a neural-network architecture that learns representations using a graph's connections together with features associated with its nodes. Graph-convolution layers combine information from neighboring nodes and apply learned transformations. The widely used Kipf-Welling formulation performs degree-normalized aggregation with self-connections, making each node's update depend on itself and its local neighbors.

What it means

An ordinary image convolution uses a regular pixel grid. A graph can have irregular neighborhoods: one node may have two neighbors and another twenty. A GCN adapts the idea of shared local computation to that structure.
The graph and node features supply different information. Features describe a node, while edges determine which other nodes contribute to its representation. The term can refer broadly to graph-convolution models, but this entry uses the standard Kipf-Welling layer as a concrete example. Other graph neural networks use different aggregation, attention, or edge-handling mechanisms.

How it works

Start with a node-feature matrix and an adjacency matrix. In the cited formulation, add self-connections and normalize adjacency using the resulting node degrees. Multiply this normalized adjacency by the current representations and a learned weight matrix, then apply an activation function.
Stacking layers allows information to pass across multiple graph hops. Training adjusts weights using an objective appropriate to the task, such as classification loss on labeled nodes. Aggregation alone is not learning: the weights and training signal matter. More layers do not guarantee better representations, and repeated mixing can make distinctions between nodes harder to preserve.

Three practical examples

A citation-network experiment predicts document categories using document features and citation-related graph structure.
A relational-data study compares a GCN with a feature-only baseline to test whether the available links add predictive value.
A graph-learning tutorial follows how one node's representation changes when a neighboring node's feature or a connecting edge changes.

Think of it this way

Think of each participant revising a worksheet after receiving a weighted summary of nearby participants' worksheets. Repeating the exchange spreads information further. The analogy is limited because the network uses defined numerical operations and learned parameters; it does not understand relationships or decide independently which messages are trustworthy.

Worked example

Take an undirected graph of two connected nodes, A and B, with scalar features 2 and 6. Add a self-connection to each node. The resulting adjacency matrix has all four entries equal to 1, and both degrees are 2. Symmetric degree normalization therefore assigns weight 1/2 to every entry.
For one illustrative GCN layer, set the learned scalar weight to 1, omit bias, and use the identity activation. A's new feature is (1/2) × 2 + (1/2) × 6 = 4. B's is also 4. The calculation shows the aggregation part of the layer with fixed demonstration parameters; it is not a trained model.
If the learned scalar weight were 3, both outputs would be 12 under the same assumptions. More importantly, the original difference between 2 and 6 has disappeared in this tiny graph after averaging. This does not prove that every GCN loses useful information immediately, but it illustrates why graph mixing and representation design deserve attention.

Illustrative case study

In a fictional archive experiment, the Cedar team predicts topic labels from document features and links. It compares a feature-only model with a GCN using a documented split and examines whether links unavailable at prediction time were accidentally included.
The team also tests a graph built from weak or irrelevant relationships. It does not assume that adding edges must improve accuracy. This invented case illustrates experimental controls and graph semantics, without reporting benchmark gains or claiming that the architecture suits every archive.

Common misconceptions and mistakes

A GCN is not simply a graph drawing passed into an image classifier. The graph's adjacency structure participates directly in the computation. It also does not require every node to have a training label in the original semi-supervised setup.
Do not assume that all graph neural networks use the same normalization. Do not treat edges as inherently helpful: they may connect dissimilar nodes or encode relations that a basic aggregation rule handles poorly. Training and evaluation must also distinguish information allowed in a transductive setting from information genuinely available for new nodes or future predictions.

Business use

A GCN is relevant when linked records carry useful structure beyond individual feature vectors. It provides a way to learn from both, but the value should be established through comparisons and appropriate evaluation rather than inferred from the presence of a graph.
For a project specification, identify node features, edge meanings, prediction targets, and the deployment scenario. Assess the cost of storing and processing the graph and the consequences of changing connections. A model that works on one fixed graph may need additional design to handle a different graph, new relation types, or changing feature distributions.

Check your understanding

Why do both nodes become 4 in the worked example?

Answer: After self-connections and normalization, each node receives half of its own scalar feature and half of its neighbor's. With weight 1 and identity activation, both compute the same average of 2 and 6.

Related terms

Graph neural network
Message passing
Semi-supervised learning
adjacency matrix
node embedding
graph convolution
degree normalization
oversmoothing

Learn next

Learn next about message passing. Add a third node connected only to B, give it a feature, and compute how the degrees and normalized contributions change. Then distinguish what changes because of topology from what changes because of the learned weight matrix.

Message passing

Sources

Kipf and Welling, original Semi-Supervised Classification with Graph Convolutional Networks paper: https://arxiv.org/pdf/1609.02907 Original paper record: https://arxiv.org/abs/1609.02907

Previous: Graph coloring Next: Graph Convolutional Policy Network