Deteksi Cyberbullying pada Twitter menggunakan Distilbert dan (Explainable Ai) Shap

Abstract

The prevalence of cyberbullying on Twitter demands a fast and transparent automated moderation system. This study designs an efficient and explainable binary cyberbullying detection system using DistilBERT and the SHAP-based Explainable AI (XAI) method. The dataset, sourced from Kaggle, consists of 13,169 raw data filtered into 12,548 clean tweets. Preprocessing was conducted without stemming and stopword removal to preserve the language's semantic context, setting a max_length of 48 tokens. The pre-trained distilbert-base-multilingual-cased model was evaluated on an 80% training and 20% testing data split, while model transparency was validated using SHAP's PartitionExplainer algorithm. Evaluation results show DistilBERT performed robustly, achieving 84% accuracy, 84% precision, 86% F1-score, and 88% recall in detecting bullying, with an Area Under Curve (AUC) score reaching 0.9207. Locally, SHAP evaluation proved that the model concentrated heavy penalty weights on abusive words like "brengsek". Globally, SHAP revealed a unique phenomenon where the most important features were dominated by neutral vocabularies such as "times. roman" and "pendidikan", which the model cleverly utilized as strong discriminative indicators to validate normal texts. In conclusion, the integration of DistilBERT and SHAP yields a classification system that is not only accurate and efficient but also transparent and logically reasoned.