Integrating Sentiment Analysis with Machine Learning for Cyberbullying Detection on Social Media

Cyberbullying using social networking sites is an extensive network issue that can have significant psychological and emotional impacts, thus, reliable detection systems play an important role. With the power of cutting-edge ML, sentiment analysis can help identify harmful interactions on a scale that is scalable, thereby creating safer online environments. This work uses the Cyberbullying Tweets dataset from Kaggle that contains labeled cyberbullying events of various categories. Pre-processing consists of normalisation, removing noise (like urls and html tags), tokenisation, WordNetLemmatizer, vaderSentiment analysis and label encoding. TF-IDF vectorisation and SMOTE oversampling methods are used for feature extraction and class imbalance, respectively. Distributions plots, word clouds etc. are used to help us appreciate the trends in bullying, and the frequency of different types of bullying. We have applied several classifiers in the suggested framework including LR, RF, XGBoost, DT, NB, SVM, ET, GB and AdaBoost with hyperparameter tuning with GridSearchCV. Evaluation based on: accuracy, precision, recall, F1-score and efficient tracking of false positives and false negatives is done using confusion matrices. The further improvements are the SMOTEENN to balance the data and a VC ensemble (MLP and Bagging combined with LR) for robust classification. Methods like LIME and SHAP, used in XAI, emphasize the most significant factors, making the AI's decisions interpretable. The Flask deployment offers real-time predictions and confidence scores with subject modelling for usability and transparency. The results show that the proposed VC algorithm outperforms all the baseline classifiers with an accuracy of 97.4% in all the evaluation parameters, and the system can be considered reliable, scalable and feasible for the detection of cyberbullying.