The rise of social media platforms such as X (formerly Twitter) and Koo has not only advanced digital communication, but also facilitated cyberbullying, leading to serious psychological and social consequences. Traditional moderation methods struggle with the scale and evolving nature of harmful content. To address this, we systematically evaluated several state-of-the-art machine learning and deep learning methods on the publicly available HateXplain dataset. Our experiments show that deep learning methods, particularly transformer-based architectures, outperform traditional machine learning approaches. We explored various word embedding techniques for different machine learning and deep learning models, including TF-IDF, ELMO, GloVe, SentencePiece, WordPiece, and Byte Pair Encoding. Among these, the BERTweet model with SentencePiece tokenization achieved the best results, reaching an accuracy of 0.7081. This slightly surpasses the highest previously reported accuracy of 0.7070 achieved using the complete HateXplain dataset in the existing literature, demonstrating the effectiveness of combining transformer-based models with suitable subword tokenization strategies for cyberbullying detection tasks.