This study aims to address the key causes of welfare blind spots—missing variables and prediction omissions—by applying a Tabular Variational AutoEncoder (TVAE)-based structural missing data imputation technique and designing a recall-centered classification model. Using welfare administrative data collected from January 2018 to November 2023, the study systematically verified the effects of variable expansion and synthetic data integration. In Phase 1, recall slightly improved by approximately 0.4–1.0 percentage points through direct variable expansion. In Phase 2, combining TVAE-imputed data led to a substantial increase in recall, from 57.7% to 94.4%. A sensitivity analysis of synthetic data combination ratios (Test 1–4) confirmed that stable predictive performance was maintained even with 10–20% synthetic data inclusion. The time-block validation further demonstrated temporal robustness, with ROC-AUC values remaining between 0.66 and 0.67 and PR-AUC between 0.31 and 0.32 across time periods. These results empirically demonstrate that TVAE-based imputation effectively enhances the quality and predictive reliability of welfare administrative data, supporting the validity of recall-oriented, data-driven decision-making in welfare policy design.