Graph Neural Networks quiz

Question 1

What is a graph and how is it represented for machine learning?

Accepted Answer

Adjacency matrix A and node feature matrix X

Answer

Only adjacency matrix

Answer

Only edge list

Answer

Degree vector only

Question 2

What is Graph Convolutional Network (GCN)?

Accepted Answer

H = σ(D̃^{-1/2} Ã D̃^{-1/2} H W) with symmetrically normalized adjacency

Answer

H = σ(A H W) without normalization

Answer

H = W X only, no adjacency

Answer

H = softmax(A) X

Question 3

What is GraphSAGE and how does it enable inductive learning?

Accepted Answer

Samples fixed neighbors and aggregates; aggregation function generalizes to unseen nodes

Answer

Requires all nodes at training time

Answer

Only works for static graphs

Answer

Uses full neighborhood without sampling

Question 4

What is over-smoothing in GNNs?

Accepted Answer

Many layers cause node representations to converge and become indistinguishable

Answer

Over-smoothing improves with more layers

Answer

Over-smoothing only occurs in GAT

Answer

Caused by low learning rate

Question 5

What is spectral vs spatial graph convolution?

Accepted Answer

Spectral: via Laplacian eigenvectors; spatial: direct neighbor aggregation

Answer

Spectral is always faster

Answer

Spatial requires eigenvectors

Answer

They are identical in practice

Question 6

What is node classification with GNNs?

Accepted Answer

Assigns labels to nodes using both node features and graph structure

Answer

Classifies entire graphs

Answer

Only uses node features, ignoring edges

Answer

Requires all nodes to be labeled

Question 7

What is GNN for molecular property prediction?

Accepted Answer

Atoms=nodes with features, bonds=edges; GNN predicts molecular properties

Answer

Treats molecules as SMILES strings only

Answer

Only predicts atomic charges

Answer

Uses 3D coordinates only

Question 8

What is heterogeneous graph and how are GNNs extended for them?

Accepted Answer

Uses separate weight matrices or attention per relation/node type

Answer

Treats all nodes identically

Answer

Only works for bipartite graphs

Answer

Requires node type embeddings added to features

Question 9

What is Graph Isomorphism Network (GIN)?

Accepted Answer

Uses MLP + sum aggregation: h_i = MLP((1+ε)h_i + Σ_j h_j)

Answer

Uses mean aggregation

Answer

Uses max pooling

Answer

Equivalent to GCN with residuals

Question 10

What is the aggregation function choice in GNNs and why does it matter?

Accepted Answer

Sum is most expressive; mean/max lose degree/multiset information

Answer

Mean is always best

Answer

Max pooling is most expressive

Answer

All aggregations are equally expressive

Question 11

What is the over-squashing problem in GNNs?

Accepted Answer

Information from large neighborhoods compressed through bottleneck edges, limiting long-range communication

Answer

Same as over-smoothing

Answer

Caused by too few layers

Answer

Solved by increasing learning rate

Question 12

What is a virtual node in GNN?

Accepted Answer

Extra node connected to all real nodes enabling global information exchange in 1 hop

Answer

A node with no features

Answer

A node with self-loops only

Answer

Only used in heterogeneous graphs

Question 13

What is a bipartite graph and how is it used in recommender systems?

Accepted Answer

Two node sets (users, items) with cross-edges; GNN propagates to capture higher-order collaborations

Answer

Has edges within each node set

Answer

Requires explicit features only

Answer

Cannot be used with GNNs

Question 14

What is the neighborhood explosion problem in GNN training?

Accepted Answer

K-hop neighborhood grows exponentially (degree^K); sampling limits to fixed size per hop

Answer

Only occurs with more than 5 layers

Answer

Solved by increasing batch size

Answer

Only a problem for heterogeneous graphs

Question 15

What is GNN for point cloud processing?

Accepted Answer

Constructs k-NN graph from 3D points and applies edge convolution for local geometry

Answer

Treats point clouds as images

Answer

Only global pooling of point features

Answer

Requires voxelization always

Question 16

What is the role of self-loops in GCN?

Accepted Answer

Ensures each node's own features are included in aggregation

Answer

Prevents gradient flow

Answer

Only needed for isolated nodes

Answer

Makes A symmetric

Question 17

What is Open Graph Benchmark (OGB)?

Accepted Answer

Large-scale standardized benchmarks with fixed splits and evaluators for fair GNN comparison

Answer

Only contains molecular graphs

Answer

Contains only 100-node graphs

Answer

Does not provide data loaders

Question 18

What is node2vec and how does it differ from GNNs?

Accepted Answer

Random walk + Skip-gram for node embeddings; transductive and feature-agnostic, unlike GNNs

Answer

Same as GCN with fewer layers

Answer

Requires labeled nodes

Answer

Always outperforms GNNs

Question 19

What is LightGCN and why remove feature transformation?

Accepted Answer

Removes feature transformation and activation; uses only normalized adjacency aggregation

Answer

Adds extra attention layers

Answer

Uses LSTM for sequence modeling

Answer

Only for bipartite graphs without GCN

Question 20

What is the difference between GCN and GAT in terms of neighbor weighting?

Accepted Answer

GCN: fixed degree-based weights; GAT: learned feature-based attention weights

Answer

GCN learns attention weights

Answer

GAT uses degree normalization

Answer

They are identical

Question 21

What is heterophily in graphs and how does it affect GNNs?

Accepted Answer

Heterophilous graphs have different-class neighbors; standard GCN aggregation hurts performance

Answer

Heterophily always improves GCN

Answer

GCN handles both equally

Answer

Homophily is rare in practice

Question 22

What is Cluster-GCN?

Accepted Answer

Partitions graph into clusters, batches within clusters to avoid neighbor explosion

Answer

Requires full graph in memory

Answer

Uses only 1-hop neighbors

Answer

Doesn't support heterogeneous graphs

Question 23

What makes GNNs permutation equivariant?

Accepted Answer

Permuting node indices permutes outputs the same way; message passing uses only local structure

Answer

GNNs are permutation invariant at node level

Answer

GNNs depend on node ordering

Answer

Permutation equivariance must be manually enforced

Question 24

What is the difference between node-level and graph-level GNN tasks?

Accepted Answer

Node-level: predict per node; graph-level: pool all nodes to predict for whole graph

Answer

They are identical

Answer

Graph-level uses only the first node's embedding

Answer

Node-level pools all nodes

Question 25

What is neighbor sampling in GNN training?

Accepted Answer

Samples fixed-size neighbor subsets per layer to bound mini-batch computation cost

Answer

Uses all neighbors always

Answer

Only for heterogeneous graphs

Answer

Samples edges not neighbors

Test yourself