Editing Graph Neural Networks (section)

== <span style="color: #FFFFFF;">Analyzing</span> ==
{| class="wikitable"
|+ GNN Architecture Comparison
! Architecture !! Aggregation !! Key Innovation !! Best For
|-
| GCN || Normalized mean || Spectral convolution approximation || Semi-supervised node classification
|-
| GraphSAGE || Sampling + mean/max/LSTM || Inductive learning on unseen nodes || Large-scale production graphs
|-
| GAT || Attention-weighted sum || Learned neighbor importance || Heterogeneous neighbor importance
|-
| GIN (Graph Isomorphism Net) || Sum || Maximally expressive (WL-test) || Graph classification
|-
| MPNN || Custom || Generalized message passing || Molecular property prediction
|}

'''Failure modes and challenges:'''
* '''Over-smoothing''' — As GNN depth increases, all node representations converge to the same value. Stacking too many layers homogenizes representations, destroying discriminative power. Mitigation: skip connections, JK-Net (jumping knowledge), layer normalization.
* '''Over-squashing''' — Information from exponentially many nodes must be compressed into fixed-size representations as layers deepen. Bottlenecks lose important long-range information.
* '''Scalability''' — Full-graph message passing requires materializing the adjacency matrix, which is infeasible for graphs with millions of nodes. Mini-batch sampling (GraphSAGE, ClusterGCN) is essential.
* '''Dynamic graphs''' — Most GNN architectures assume static graphs. Real-world graphs (social networks, transaction graphs) evolve over time. Temporal GNNs (TGNN, EvolveGCN) address this.
* '''Heterogeneous graphs''' — Many real-world graphs have multiple node types and edge types. Standard GNNs treat all the same; HGT (Heterogeneous Graph Transformer) handles this.
</div>

<div style="background-color: #483D8B; color: #FFFFFF; padding: 20px; border-radius: 8px; margin-bottom: 15px;">