To attain optimal performance in text classification tasks, preprocessing is a widely acknowledged prerequisite. Our study aims to investigate the extent to which text preprocessing methods and tools affect cyberbullying classification in the social media domain specifically for Malay, a resource-poor language. Using a TikTok corpus containing 5000 Malay messages, we explore nine text preprocessing methods, stemming tools (Naïve, HuggingFace and Sastrawi stemmers) and word representations (BoW, TF-IDF, Word2Vec and Mistral) using support vector machine (SVM) in four different cyberbullying classification tasks (cyber-aggression classification, cyber-aggression type classification, cyberbullying classification and cyberbullying role classification). Results show that our proposed nine text preprocessing methods generally produce better outcomes than classifiers without preprocessing. Our ablation experiments further pinpoint stemming, stopword removal and text normalization play a significant role in text preprocessing while special token removal and word correction yield only marginal effect. Thus far, our study is the most comprehensive attempt to examine the effect of Malay text preprocessing methods across not only multiple word representations but also the generalizability across multiple classification tasks.