Answers are grounded in this paper and cite their sources.
Graph-regularized count estimators are usually evaluated by prediction loss, leaving open how revisions to the counts or graph affect activity concentration and proportional budgets. We derive deterministic log-rate perturbation bounds for ridge-regularized Poisson estimation with a fractional graph Laplacian and propagate them to top-decile shares and budget reallocation. In three chronological evaluations using a KuaiLive-M3 release with 213,027 successful shares, the independent maximum-likelihood estimator gives lower test deviance than the graph fits. The ordering persists at zero ridge across the ordinary and polynomial fractional fits and in 13 rolling refits of the ordinary graph. Matched fractional simulations use 120-node paths. In the concentrated design at λ = 10, α = 0.35 minimizes the mean squared log-rate error and budget reallocation, and α = 0.60 minimizes the mean top-decile share error. Independently tuned results depend on the signal and loss. The allocation results quantify the sensitivity of a specified rule. Its economic value remains to be evaluated with observed resource assignments and outcomes.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: