This study addresses data imbalance in healthcare and finance through a novel domain-adaptive Conditional GAN framework with theoretically grounded, specialized loss functions and adaptive weight adjustment mechanisms. We introduce domain-specific loss functions based on established domain knowledge: feature correlation preservation for medical diagnosis and temporal consistency for financial fraud detection. The adaptive weight adjustment mechanism employs mathematically rigorous formulations with proven convergence guarantees under bounded gradient conditions. Comprehensive empirical validation on Wisconsin Breast Cancer and Credit Card Fraud datasets demonstrates significant improvements over state-of-the-art methods. Our approach achieves Feature Correlation Preservation scores of 0.849±0.004 versus 0.000 for traditional methods, and superior F1-score performance with improvements over recent competing methods while maintaining faster training times. The method provides exceptional robustness across extreme imbalance ratios from 1:10 to 1:2000, with advantages increasing at higher imbalance levels. Statistical validation using 5-fold stratified cross-validation with Bonferroni correction confirms significance (p 0.001) with large effect sizes (Cohen's d 1.4). Scalability analysis demonstrates practical applicability up to 1 M samples with manageable computational overhead. This work establishes a foundational framework for domain-aware synthetic data generation with proven theoretical guarantees and broad implications for imbalanced learning across critical application domains.
Yoo et al. (2026) studied this question.