Feature Drift‐Aware Boosted Ensemble (FDABE): A Dynamic Feature Weighting Framework for Concept Drift Adaptation in E‐Commerce Data Streams

ABSTRACT Concept drift is a central challenge in e‐commerce stream analytics because customer preferences and interaction patterns evolve over time. This paper proposes the Feature Drift‐Aware Boosted Ensemble (FDABE), which adapts at the feature level before updating the boosted learner. Its Feature Drift Severity Index (FDSI) combines Jensen‐Shannon distributional change, exponentially decayed drift memory, and temporal variation in feature relevance to drive dynamic feature influence and drift‐regularized boosting. On three public e‐commerce benchmarks, FDABE achieved 96.2% accuracy, 95.4% F1‐score, and 97.0% ROC‐AUC, with a 72‐sample detection delay and 148‐sample recovery time. A five‐seed confirmatory rerun across abrupt, gradual, incremental, recurring, subtle‐sparse, and noise‐obscured drift evaluated Precision, Recall, F1‐score, Balanced Accuracy, and PR‐AUC. FDABE significantly outperformed Static XGBoost, Incremental Weighted Ensemble, and Dynamic Ensemble Selection, while gains over Online Feature Screening were smaller. Differences versus Adaptive XGBoost, Evolving Gradient Boost, and JSD‐only Dynamic Feature Weighting were not significant after Holm correction. These results support FDABE as a competitive feature‐aware adaptation mechanism without claiming universal superiority.