Social networking platforms give users countless opportunities to share information, collaborate, and communicate positively. The same platform can be extended to a fabricated and poisonous atmosphere that gives an impersonal, harmful platform for online misuse and assault. Cyberstalking is when someone uses an internet system to ridicule, torment, insult, criticize, slander, and discredit a victim while never seeing them. With the growth of social networks, Facebook has become the online arena for bullying. Since the effects could result in a widespread contagion, it is vital to have models and mechanisms in place for the automatic identification and removal of internet cyberbullying data. This paper presents a robust hybrid ML model for cyberbullying detection in the Bengali language on social media. The Bengalibullying proposal involves an effective text preprocessing to make the Bengali text data into a useful text format, feature extraction using the TfidfVectorizer (TFID) to get the beneficial information of text data and resampling by Instance Hardness Threshold (IHT) procedure to balance the dataset to avoid overfitting or underfitting problems. In our experiment, we used the publicly available Bangla text dataset (44,001 comments) and got the highest performance ever published works on it. The model achieved the most elevated accuracy rate of 98.57% and 98.82% in binary and multilabel classification to detect cyberbullying on social media in the Bengali language. Our best performance findings are more effective than any previous effort in identifying and categorizing bullying in the Bengali language. As a result, we might use our model to correctly classify Bengali bullying in online bullying detection systems, protecting people from being the targets of social bullying.