Semantic Feature Engineering with LSA-SVM for Cyberbullying Comment Classification on Instagram

Social media is now an essential part of everyday life, with Instagram being one of the most popular platforms and often utilized for various purposes, one of which is to increase popularity. However, the platform also often becomes a place where acts of violence and impoliteness in commenting increase, known as cyberbullying. To address the problem, detecting and classifying cyberbullying comments on Instagram is an important step in cyberbullying prevention. However, in text classification, several challenges need to be overcome to ensure the success of the model, such as polysemy, curse of dimensionality, and selection of text representation for feature extraction. Therefore, this study aims to implement a feature engineering technique using a hybrid approach that combines word weighting with TF-IDF and LSA method to reduce feature dimensionality and capture the semantic meaning of the data, with SVM used as a classifier to classify bullying and non-bullying comments. The results showed that the proposed method using feature engineering of the LSA matrix formed from the dataset of one of the classes, yielded a significant accuracy of 98%. In comparison, the conventional method with feature engineering using TF-IDF and the use of LSA matrix formed from the dataset of both classes only achieved an accuracy of 84%. This shows that the proposed method is more effective than the baseline approach.