Sequence-Aware GAN-Based Augmentation for Leakage-Safe Next-Session Payment Prediction under Extreme Class Imbalance

Extreme class imbalance remains a major obstacle in behavioral prediction tasks where rare positive events carry disproportionate practical value. This challenge is particularly severe in free-to-play environments, where payment events are sparse, sequentially conditioned, and vulnerable to label leakage if prediction targets are not formulated carefully. In this study, we propose a sequence-aware data augmentation framework for leakage-safe next-session payment prediction in highly imbalanced session-based behavioral data. Unlike standard tabular GAN oversampling, which generates isolated minority records, our padding-aware GAN-based behavioral-window synthesis models consecutive pre-payment sessions, preserving recent engagement and progression context relevant to monetization prediction. The resulting synthetic trajectories augment the training data within a unified classification pipeline. We evaluate the proposed approach against random oversampling, SMOTE, class weighting, threshold optimization, and their combinations using a fixed neural prediction backbone. Experiments on a real-world production dataset containing tens of millions of sessions with an extreme positive rate of approximately 0.34% show that conventional oversampling methods provide limited practical benefit, whereas GAN-based augmentation consistently improves minority-class ranking. In the main player-session experiment, the GAN-based model increased PR-AUC from 0.0044 to 0.0140, corresponding to an approximately 3.18-fold lift over the baseline, while its threshold-calibrated variant achieved the highest F1-score among the evaluated player-session scenarios. Auxiliary experiments on the Spambase benchmark provide supporting evidence that the pipeline can also improve minority-class modeling in a static tabular setting. The results highlight the importance of behaviorally coherent minority synthesis and precision-recall-oriented evaluation.