Match graph evaluation splits to the claimed generalization

In a transductive node-classification protocol, unlabeled test-node features and graph structure can be available during training while test labels remain hidden. That is legitimate when declared, but it does not establish generalization to entirely unseen nodes. In a strict node-inductive protocol, held-out test nodes are excluded from the training computation graph, including neighborhood sampling and fitted feature statistics. At inference, their available features and permitted neighborhoods can be used according to the deployment scenario. Inductive capability still depends on meaningful features and distributional coverage.

GraphSAGE learns feature-and-neighborhood aggregation rather than requiring a trained vector for every node ID, enabling its studied unseen-node setting. A model can lose that practical capability if its pipeline adds a required learned ID table without an unseen-node policy. Entirely unseen graphs pose an additional shift in topology, features and label distribution. Split by entity, source graph or time when that matches deployment, and prevent related duplicates or inverse edge records from crossing boundaries in a way that reveals target information.

A chronological edge split does not prevent leakage if neighborhood features or degrees were computed from the final graph. Enforce event and availability cutoffs for every derived input, and specify whether earlier validation or test events update state after being scored. Hold the candidate set and negative sampling protocol fixed when comparing models. Easy random nonedges and historically plausible candidates can yield different ranking difficulty, so their metrics are not interchangeable estimates of one unnamed task.

Average precision rewards positive items near the top of a ranked candidate list, averaging precision at each positive’s rank. Candidate composition matters because high-scoring negatives push those ranks downward. State the number and distribution of candidates, tie handling and whether metrics are computed per source then averaged or pooled globally. Precision at a random draw equals candidate prevalence, but that observation is not an exact finite-sample formula for expected average precision. Avoid using a balanced sampled classification metric as a deployment retrieval guarantee.

Nodes in the same connected region can share features, neighbors and labels, so treating all test nodes as independent can understate uncertainty. Use a design aligned with the intended generalization unit, such as held-out graphs, entities or temporal blocks, and report critical slices by degree, node type and novelty. Repeat stochastic training and sampling under a declared seed design. Training-seed variability is different from uncertainty about future deployment graphs, and neither is eliminated by a large count of highly related test edges.

For an independent-sample comparison, Reset the sampling lab and set both p=0.2 and n=25. Predict variance 0.0064 and standard deviation 0.08 for the sample mean. Separately Reset and set both p=0.2 and n=100; predict variance 0.0016 and standard deviation 0.04. Play highlights exact binomial outcomes. This iid Bernoulli instrument illustrates a sample-size law only; it does not compute AP, capture graph dependence or justify counting connected evaluation nodes as independent.