Week 13: Graph Neural Networks

Question 1

Shortly describe what is a GNN?

Answer 1

A GNN is an optimizable transformation on all attributes of the graph (nodes, edges, global-context) that preserves graph symmetries (permutation invariances)

Question 2

What are the three main types of prediction tasks in GNNs?

Answer 2

Graph-level: Predict properties of entire graph (e.g., “Does molecule smell?”)
Node-level: Predict properties of individual nodes (e.g., Karate club allegiances)
Edge-level: Predict properties of edges or edge existence (e.g., relationship classification)

Question 3

What are the three geometric priors of the GDL blueprint?

Answer 3

A:

Symmetry (Invariance/Equivariance): Output should be consistent under transformations
Geometric Stability: Stability to signal and domain deformations
Scale Separation: Multi-scale representation capability

Question 4

Explain the mechanism of message passing in graph neural networks.

Answer 4

A: Message passing works in three steps:

Gather: Collect neighboring node embeddings (messages) for each node

Aggregate: Combine messages using aggregation function (sum, mean, max)

Update: Pass aggregated messages through learned neural network to update node representations

Question 5

What are the two main approaches to graph convolution?

Answer 5

A:

Spectral-based:

Filter signals using eigendecompositions of graph Laplacian
Transform to spectral domain, filter, then inverse transform
Removes noise from graph signals

Spatial-based:

Filter by information propagation distance
Aggregate based on neighborhood structure
Direct aggregation of neighboring features

Question 6

What made AlphaFold2 revolutionary?

Answer 6

A:

Solved 50-year protein folding challenge
Achieved breakthrough performance on CASP competition
Created comprehensive protein structure database
Won 2024 Nobel Prize in Chemistry
Enables drug discovery, understanding life processes