Cyberbullying on Bengali social media platforms has grown into a serious societal concern, yet automated detection remains challenging due to the morphological richness and low-resource nature of the Bengali language. Existing studies have largely relied on binary classification, single-platform corpora, or computationally expensive transformer architectures, leaving a clear gap in robust multiclass detection frameworks that balance accuracy with practical deployability. There is a pressing need for a systematic approach that exploits the complementary strengths of classical and neural models within a unified pipeline. This study proposes a hybrid ensemble framework combining Random Forest, Multi-Layer Perceptron, Convolutional Neural Network, and Recurrent Neural Network as base learners, integrated through soft voting, hard voting, and a CNN-based stacking meta-learner. Experiments were conducted on a publicly available Bengali cyberbullying dataset of 6,005 samples, augmented to 12,500 balanced instances across five abuse categories— Political, Troll, Sexual, Threat, and Neutral. A comparative evaluation of fourteen models demonstrates that the standalone CNN achieves 92.6% accuracy with a ROC AUC of 0.989, outperforming prior Bengali cyberbullying detection baselines by approximately 2.8 percentage points, whilst the stacking ensemble attains the highest discriminative power at 0.990 AUC, establishing a reproducible benchmark for fine-grained Bengali abuse detection.

Bengali Cyberbullying Detection Hybrid Ensemble Learning Convolutional Neural Network Multi-Class Text Classification Natural Language Processing Low-Resource Language Processing