Define what a negative edge means before sampling it

A sampled pair can be a known positive accidentally retained by an approximate sampler, an unobserved true relation, or a pair that becomes positive during the target horizon. Those are different reasons a proposed negative may be inappropriate. Define candidate node types, self-pair rules, edge direction and prediction horizon first. A future interaction is not automatically a false negative for a task whose label is explicitly “no edge yet”; it is problematic when it contradicts the declared horizon label or asserted permanent absence.

For offline training, matured labels within the allowed training information boundary can legitimately define positive and negative outcomes. But filtering candidates using edges beyond that boundary can grant future knowledge. Keep label construction and feature availability separate: observing a future outcome for supervision does not permit its edge or derived features to appear in the earlier message graph. Treat censored horizons carefully when the observation period ends before a candidate’s full outcome window; missing follow-up is not reliable proof of a negative label.

With independent uniform proposals and rejection of those known positives, the expected proposals per accepted draw are the reciprocal of the acceptance probability. Dense neighborhoods can therefore make rejection expensive. An implementation can sample directly from a known complement, but constructing or maintaining that complement has its own cost. If proposals are nonuniform or sampling is without replacement, rederive the cost rather than importing the same geometric calculation. Report duplicate negative pairs and whether their multiplicities affect the loss.

Uniform random negatives can be much easier than plausible alternatives such as historically observed or semantically similar pairs. Changing the negative distribution changes the training contrast and evaluation candidate difficulty. The dynamic link-prediction evaluation literature explicitly demonstrates the importance of that protocol. If importance correction is needed for a declared target distribution, use the actual proposal probabilities and support; no weighting scheme recovers candidates that the proposal can never draw. A higher metric after making negatives easier is not evidence of a better graph model.

PyTorch Geometric 2.7.0 LinkNeighborLoader warns that its approximate negative sampling can include false negatives and that supervision edges are not automatically masked out of message passing. Removing a target from a label list is therefore insufficient if the same edge remains in the sampled topology. For undirected relationships, also consider the reverse record; for repeated interactions, choose the correct time-specific exclusion. Preserve legitimate historical edges when the task allows them, while excluding the target information that would make the prediction circular.

For a contamination analogy, Reset the probability lab and set both p=0.2 and n=10. Predict two problematic draws on average and probability of none 0.8¹⁰≈0.1074. Play highlights exact binomial outcomes rather than sampling graph pairs. This assumes independent equal-probability contamination and does not inspect an adjacency structure, detect unknown positives or enforce supervision-edge masking. Its useful lesson is that even modest per-draw contamination can affect many minibatches; actual prevalence and dependence require evidence.