Graph Neural Networks Quiz
Five questions spanning the course — the message-passing framework, GCN and GraphSAGE, graph attention, readout, and over-smoothing.
Exercise
Which property of sum/mean/max aggregation makes it the right choice for combining a node's neighbors?
Need a hint?
Nodes have no canonical order.
Exercise
In the GCN layer H' = σ(Â H W), why is the adjacency augmented with self-loops (Â built from A + I)?
Need a hint?
What would a node's update miss without its own row?
Exercise
What two problems with a vanilla GCN does GraphSAGE address?
Need a hint?
Think memory on huge graphs, and brand-new nodes.
Exercise
A GAT's attention coefficients α_vu for node v are normalized with a softmax. Over what set is that softmax computed?
Need a hint?
GAT attention is local, not global.
Exercise
Which technique does NOT help mitigate over-smoothing in deep GNNs?
Need a hint?
Three of these preserve node distinctiveness; one makes the problem worse.
Exercise
A GNN aggregates neighbors with a symmetric function (sum/mean/max) so relabeling nodes doesn't change the output. Which slot of the ML project loop is that design, and what does it encode?
Need a hint?
It's a structural assumption about the model — an inductive bias for graphs.
Exercise
Transfer test — a technique this course did NOT teach: an A/B test comparing a new recommendation model against the current one on live traffic. Which slot of the ML project loop does it belong to?
Need a hint?
It measures real impact after deployment and feeds the decision to ship.