> ML_LITERATURE // KIPF-WELLING-2016-SEMI-SUPERVISED-CLASSIFICATION-GRAPH-CONVOLUTIONAL-NETWORKS_v1.0
Semi-Supervised Classification with Graph Convolutional Networks (GCN)
Thomas N. Kipf, Max Welling · International Conference on Learning Representations (ICLR) (2016)
seminal-architecture2016foundationalthirdPartyReproduced
Principal Contribution
Derived a localized first-order spectral filter approximation for semi-supervised node classification on graph-structured data.
Operational Relevance
The foundational benchmark model in PyTorch Geometric and DGL for social network node classification and molecule property prediction.
Assumptions
- Homophily assumption: graph structure encodes similarity, and self-loops with symmetric normalized adjacency propagate features smoothly
Limitations
- Full-batch training requires keeping the full adjacency matrix in GPU memory; over-smoothing when depth exceeds 4 layers