Skip to the content.

Graph Neural Networks (GNNs)

Graph Neural Networks are a class of deep learning models designed to operate on graph-structured data — nodes (vertices) connected by edges (relationships).

Core Mechanism: Message Passing

GNNs follow a message passing (neighborhood aggregation) paradigm:

  1. Each node starts with a feature vector
  2. Aggregate: collect features from neighbors
  3. Update: combine current features with aggregated neighbor information
  4. Repeat for K layers — each node incorporates information from K-hop neighbors
h_v^(k) = UPDATE( h_v^(k-1), AGGREGATE({ h_u^(k-1) : u ∈ N(v) }) )

Key Architectures

Architecture Key Idea Strengths
GCN Spectral convolution with mean aggregation Simple, effective baseline
GAT Attention-weighted neighbor aggregation Differentiates neighbor importance
GraphSAGE Sampling + inductive aggregation Scales to large/dynamic graphs
GIN Sum aggregation, maximally expressive under WL test Strongest theoretical guarantees
MPNN General message passing framework Unifying abstraction

Common Tasks

Strengths and Limitations

Strengths: handles relational data naturally, permutation invariant, captures local and global structure.

Limitations: over-smoothing with many layers, scalability challenges on billion-node graphs, expressiveness bounded by the Weisfeiler-Leman test.