Build graph state from information available at prediction time

Let an event have an occurrence timestamp and a timestamp when its usable record became available. For a query at time q, this lesson admits events strictly earlier than q that were available no later than q. A different application can choose an inclusive event-time boundary, but it must resolve simultaneous events consistently. A timestamp alone may not order events with equal times, so use a declared tie key or a simultaneous-batch rule. Features and corrections need the same availability discipline as edges.

An offline warehouse can contain late records whose event times precede a historical query even though the records arrived afterward. Filtering only event time would grant the historical predictor knowledge it lacked. Conversely, a very recent available event can still be disallowed by a task’s strict event-time rule. PyTorch Geometric 2.7.0 NeighborLoader documents a timestamp constraint with earlier-or-equal neighbors; that API condition alone does not establish an application’s ingestion-time availability or implement the stricter boundary adopted here.

A temporal encoder can maintain node memory updated by prior interactions. When predicting an interaction, compute its score from the eligible prior state and only then update memory with the observed interaction under the chosen protocol. Updating first can expose the target itself. Temporal Graph Networks make memory and event-driven updates explicit, but their architecture does not excuse a pipeline from enforcing the evaluation’s information boundary. Batching several events together also needs a policy preventing later batch events from influencing earlier predictions.

Late arrivals pose a state-reconstruction question. An additive state can sometimes incorporate a delayed event directly; an order-sensitive recurrent update may require replay from a checkpoint in event order or a deliberately different arrival-order model. Replaying with newly arrived history can improve a current prediction but must not overwrite what an earlier prediction was allowed to see. Preserve state or event-log versions for reproducibility. Two requests for the same node at different cutoffs can therefore require different representations and cannot be merged solely by node ID.

Version the event log, availability timestamps, feature snapshots, state initialization and ordering policy. Report whether validation or test events update memory after being scored; online adaptation and a frozen-state evaluation answer different questions. Ensure negative candidates and derived degree features obey the same cutoff. A future-computed aggregate can leak even when the raw target edge is removed. Audit small timelines by hand before trusting a large temporal sampler, including late arrivals, tied timestamps and repeated interactions between the same nodes.

The Markov lab illustrates ordered state transitions in a deliberately simple family. Reset and set both p=1 and q=1. Predict probability one in state 0 initially, probability one in state 1 after one Step, then probability one in state 0 after a second Step. Each Step advances one transition. This exact alternating chain is neither a sampled event stream nor a temporal GNN. It makes before-update and after-update states visibly different without claiming to enforce real event-time or availability rules.