Why Graph-Structured Data?
Much of the world’s most interesting data comes with relational structure baked in: users and their social connections, atoms and their molecular bonds, entities and their links in a knowledge base. Standard deep learning architectures — CNNs, RNNs, Transformers — excel at grid or sequence data, but they don’t have a native way to handle the irregular topology of a graph.
Graph Neural Networks (GNNs) address this gap by providing a unified framework for learning directly over graph-structured data.
The Message Passing Paradigm
Nearly every modern GNN architecture can be described through the lens of message passing:
- Aggregate: each node collects feature messages from its neighbors
- Update: the node combines aggregated neighbor information with its own features to produce a new representation
Formally, the layer-$l$ update for node $v$ is:
h_v^(l) = UPDATE(h_v^(l-1), AGGREGATE({h_u^(l-1) : u ∈ N(v)}))
The differences between architectures boil down to design choices in AGGREGATE and UPDATE.
A Tour of the Major Architectures
GCN (Graph Convolutional Network)
Kipf & Welling (2017) grounded GCN in spectral graph theory, using symmetric normalization when aggregating neighbors:
- Strengths: simple to implement, computationally efficient
- Limitations: all neighbors receive equal weight; no selectivity
GAT (Graph Attention Network)
Veličković et al. (2018) introduced attention to GNNs, dynamically computing per-edge aggregation weights:
- Strengths: learned edge importance, better interpretability
- Limitations: higher compute cost, more hyperparameters
GraphSAGE
Hamilton et al. (2017) proposed sampling a fixed-size neighborhood at each layer, making GNNs tractable on very large graphs:
- Strengths: inductive learning, scales to millions of nodes
- Limitations: sampling introduces variance; training can be less stable
Where Things Stand
[TODO] Add specific experimental findings or application examples here.
Graph learning remains a rapidly moving field — from the early GCN days through the current wave of Graph Transformers that marry attention with topology. The lab’s focus within this space is on… (please fill in).