Drift-Aware Dynamic Ensemble Learning for Adaptive Email Spam Detection

Email spam evolves every day with disguised text, changing campaigns, dangerous links and quick changing message patterns. This research aims at proposing a different email spam detection ensemble-learning framework from the one proposed in the IEEE MENACOMM-2025 project which combines SVC, MLP, and ET. The proposed system uses character n-gram feature hashing and light-weight message statistics, and consists of three online linear base-learners: Ridge Classifier, SGD Classifier with modifiedHuber loss, Passive-Aggressive Classifier. DWM ensemble: an ensemble changes the weight of the learners over time and ADWIN detects changes in the distribution of emails and adapts the weight of the weak learners. The system is deployed with a Django based email client interface which takes the fields of sender, topic and message content and returns a Spam or Legitimate/NonSpam with confidence and key indicators. The comparison illustrates that the proposed DWM+ADWIN configuration has improved accuracy of 98.2%, precision of 97.7%, recall of 98.0% and F1-score of 97.8%, and recovers faster in a temporal distribution shift, compared to static classifiers. The focus of the framework is on adaptation, low memory text representation, operational false positive control, and practical web deployment.