The Impact of TVAE-Based Synthetic Data Utilization on Feature Importance in Welfare Blind-Spot Prediction Models: A Comparative Analysis of Original and Combined Datasets

Welfare blind spots arise when households in need are not identified in a timely manner, partly because policy indicators are introduced gradually and are therefore structurally missing in earlier administrative records. This study proposes a recall-oriented decision-support framework that integrates a Tabular Variational Autoencoder (TVAE) to reconstruct historically unavailable policy variables and to examine how progressive feature availability affects model performance and interpretability. Using 2,630,195 welfare-crisis records collected between 2019 and 2023, we designed a four-stage empirical procedure. In Stage 1, a complete-data benchmark was established using the most recent timeline to evaluate whether incremental feature expansion improves classification performance without compromising recall. Stage 2 applied a time-split training strategy to prevent temporal information leakage when generating synthetic values for late-introduced policy variables, and the fidelity of the generated data was assessed through distributional similarity measures. Stage 3 examined recall-oriented classification under progressive feature augmentation using tree-based ensemble models with threshold tuning aligned with policy objectives. The results show that expanding feature availability improved overall discrimination while maintaining high recall suitable for omission-minimizing screening. In Stage 4, feature importance was compared across cumulatively combined datasets to trace how key risk factors evolved as historical coverage expanded. The analysis identified risk factors whose influence remained stable across data expansion, as well as factors whose importance emerged only after additional policy variables became available. Overall, the findings demonstrate that leakage-controlled generative reconstruction combined with recall-focused decision boundaries can enhance the robustness and interpretability of welfare screening systems operating under structurally missing policy information.