Dart: Adaptive Tweedie Objective Mitigates the σ Escape Route in Cross-Platform Popularity Prediction
TL;DR
An adaptive Tweedie objective is studied—unit deviance with dispersion fixed at ϕ = 1—which couples residual weighting to the mean (Var(Y) = ϕμp) and matches the support of the targets and is a first-class design choice.
Abstract
Social-media popularity prediction is hampered by a mismatch between its targets—non-negative, zero-inflated, and heavy-tailed—and the Gaussian negative log-likelihood (NLL) and mean squared error (MSE) objectives commonly used to fit them. We examine one consequence, for which we use the descriptive shorthand σ escape route: under heteroscedastic Gaussian NLL an independently predicted scale absorbs large residuals and attenuates the gradient for the mean μ. As a distribution-aware alternative we study an adaptive Tweedie objective—unit deviance with dispersion fixed at ϕ = 1—which couples residual weighting to the mean (Var(Y) = ϕμp) and matches the support of the targets. Architecture-controlled swaps improve Spearman’s ρ for every predictor tested, and Dart, a retrieval-augmented instantiation assembled from established components, attains the highest average ρ across nine platforms in both per-platform (0.395) and leave-one-platform-out (0.275) evaluation, without using any target-platform data. A conventional log (1 + y)-MSE control nevertheless remains competitive under LOPO and is stronger on several top-k metrics, and retrieval helps unevenly across platforms. Objective selection is therefore a first-class design choice, separable from and complementary to retrieval-based context.