When Can You Correct Distribution Drift in Temporal Graph Generation? A Sharpening–Drift Tension and an Impossibility for Observation-Based Correction

Generative models for temporal graphs are typically trained on a specific historical period and deployed on a subsequent one. This paper demonstrates that such models suffer from a fundamental degradation in performance due to distribution drift. The authors show that the masked flow-matching loss decomposes into an irreducible entropy term and a divergence term that increases as the model 'sharpens' its predictions on the training data. Specifically, the model becomes increasingly confident in structures that were rare during training but common during deployment, leading to an error floor that cannot be mitigated by increasing the sampling budget.

Because the deployment period is observable, it is tempting to assume that one could measure the drift and re-aim the model accordingly. The authors prove that this is information-theoretically impossible for observation-based correctors. They demonstrate that the target statistics behave like a fast mean-reverting random walk where the one-step innovation is as large as the drift itself. Consequently, any corrector relying on past observations is limited by the conditional variance of the target statistic, and trend extrapolation performs worse than simply using the last observed state.

Across nine distinct domains and multiple token resolutions, the authors observe that the drift-period marginal error remains largely invariant to the sampling budget, while the error floor is significantly higher than the in-period floor. An oracle that has access to the current period's statistics can remove roughly 60% of the error, whereas the best possible observation-based corrector recovers only about 5.7%. The authors conclude that effort should not be directed toward better estimators of the observed past, but rather toward incorporating side information or developing models that avoid over-sharpening.