Двометричний протокол оцінки точності розподілу при імпутації відсутніх даних

Abstract

This paper empirically tests the «RMSE illusion» in missing data imputation: the finding that deterministic iterative imputers (MICE, MissForest) minimize reconstruction error while causing Distributional Collapse – measurable degradation of marginal variance and covariance structure. Using the ACI dataset under MNAR conditions, we compare deterministic and stochastic imputers via a dual-metric protocol covering both reconstruction accuracy (RMSE, MAE) and distributional fidelity (Wasserstein distance, VR, KS). Results confirm a fundamental trade-off: deterministic methods achieve low RMSE but fail distributional preservation, while stochastic approaches do the opposite. This motivates a hybrid design combining deterministic initialization with stochastic refinement.