We identify a Minority Signal Dilution (MSD) effect in hybrid GAN oversampling: when CTGAN is trained on a SMOTE-balanced mixture (A3: SMOTE→CTGAN), majority-class structure dilutes the generator learning signal for fault patterns. Training CTGAN exclusively on minority samples (A4: minority-only CTGAN→SMOTE) eliminates MSD by construction. A controlled ablation study across five real-world industrial datasets (IR: 16:1–578:1) shows that A4 outperforms A3 by up to 3.2 AUC pp (p<0.05) under high imbalance and low Sample-to-Dimension Ratio (SDR<13.8). Spearman ρ=−0.504 (p=0.033) between SDR and A4 AUC advantage across 18 datasets. A two-step decision framework (IR screen → SDR screen) is derived for practitioners. Submitted to Computers in Industry (Elsevier).
Chao et al. (Sun,) studied this question.