Wave-GA: Cyclical Evolutionary Optimization for Streaming Classification with Emerging Labels

Streaming classification systems must maintain robustness under the "holy trinity" of non-stationarity: concept drift, extreme class imbalance, and evolving label spaces (class-incremental learning). In these regimes, standard online learners and ensemble methods frequently suffer from catastrophic forgetting of minority concepts or lack mechanisms to accommodate newly emerging classes without retraining. In this paper, we introduce Wave-GA, an evolutionary framework that evolves a transparent linear model through cyclical, class-focused training waves. By rotating optimization pressure across classes and utilizing dynamic matrix expansion, Wave-GA integrates distinct mechanisms for drift adaptation (per-class replay buffers), imbalance handling (minority-aware blended fitness), and emerging labels (architectural growth). Extensive prequential evaluation on six synthetic benchmarks and three real-world streams demonstrates that Wave-GA prevents the minority-class collapse observed in state-of-the-art ensembles like Adaptive Random Forest. The approach improves average balanced accuracy by 6–12 percentage points over standard evolutionary baselines and delivers robust long-horizon stability with constant-time inference.