Practice — Graph ML & Graph Neural Networks (6 questions)

A mobile operator has 2,000,000 subscribers, a monthly churn rate of 3 percent, and a call graph (who calls whom, with call counts and timestamps). A data scientist proposes "a GraphSAGE model on the call graph" to predict next-month churn, because churn is known to spread through friend groups. The current model is a LightGBM on billing and usage features.

1. Before any model, how would you check the claim that churn spreads along the call graph? Compute the edge homophily expected by chance at a 3 percent churn rate and explain why the raw edge homophily is the wrong statistic to report.
2. Propose a baseline ladder from the current model to a GNN, naming the graph features at each rung.
3. Name two leakage traps in a feature such as "share of this subscriber's contacts who have already churned".
4. State the decision rule you would use to accept or reject the GNN.

You have 120 days of events from a two-sided marketplace graph: buyer and seller accounts as nodes, transactions as timestamped edges. Accounts are labelled "fraudulent" by a review process and chargebacks that mature 30 days after the account's first transaction. New accounts join every day. A colleague has already trained a GNN with a random 80/20 split of nodes and reports PR-AUC 0.90.

1. Explain three distinct ways the random node split leaks.
2. Design the split: snapshot days, which nodes are labelled in train, validation and test, and which edges each model may see.
3. State what "inductive" requires of the model and which of (a) free per-node embeddings, (b) GCN fit on one fixed adjacency, (c) GraphSAGE with attribute inputs, (d) a GBM on as-of graph features qualify.
4. List the slices you would report in addition to the pooled metric.