Social media platforms play a crucial role in the dissemination of cyberbullying, making automated detection challenging due to the varied forms of content. This task is further complicated by class imbalance, which leads to biased machine learning models. We propose a hybrid solution for detecting cyberbullying using imbalanced datasets. First, to address data imbalance, we employ techniques such as Random Oversampling, SMOTE (Synthetic Minority Oversampling Technique), and Borderline-SMOTE. Second, to capture essential features and patterns while reducing noise, we introduce a Bi-LSTM-based Auto-encoder (AE). Experimental results using standard evaluation metrics demonstrate the efficacy of our proposed method for cyberbullying detection, achieving F1-scores of 0.71, 0.89, and 0.72 across three benchmark datasets. These results represent a significant improvement over existing approaches, particularly in addressing class imbalance.