GCN (Graph Convolutional Networks)

Graph Convolutional Networks (GCNs) extend the core idea of convolutional neural networks to non-Euclidean, graph-structured data. Traditional CNNs exploit the regular grid structure of images by sliding a filter across spatial neighborhoods. GCNs generalize this operation to irregular graphs by aggregating feature information from a node's local neighborhood. The mechanism multiplies node feature matrices with a normalized version of the graph's adjacency matrix, allowing each node to collect and transform signals from its directly connected neighbors. Stacking multiple such layers propagates information across increasingly distant parts of the graph, producing context-aware node representations.

A GCN layer involves three steps: aggregating neighbor features, linearly transforming the result with learned weight matrices, and applying a nonlinear activation function. Normalization, typically by node degree, prevents high-degree nodes from dominating the aggregation. This spectral interpretation draws on graph signal processing. Kipf and Welling's 2016 simplification made it computationally tractable by approximating expensive spectral graph convolutions with an efficient first-order localized filter. The resulting model scales to large graphs and is straightforward to implement and train.

GCNs are foundational in graph machine learning, with strong results on node classification, link prediction, and graph-level classification. Applications include molecular property prediction in drug discovery, fraud detection in financial networks, knowledge graph reasoning, and recommendation systems where users and items form a bipartite graph. GCNs also led to a broader family of graph neural network architectures, including GraphSAGE, GAT, and GIN, each refining how neighborhood information is aggregated or weighted. GCNs remain essential for work with relational or structured data.