K-Nearest Neighbors Undersampling as Balancing Data for Cyber Troll Detection

Cyber Troll is a behavior to provoke and attack one's emotion online. It happens throughout the world, including Indonesia. One of the social media platforms in which cyber trolling frequently happen is Twitter. There are several ways to deal with cyber trolling, one of which is to implement Support Vector Machine Classifier to separate between cyber troll tweets and normal tweets. Since most of the tweets are normal tweets, it is hard to obtain a balance between them and this leaves us at 1:2 ratio or even more. Just like most machine learning algorithm, SVM can show poor performance on the minority class because SVM was designed to induce a model based on the overall error. In this paper, we tried to implement KNN-Undersampling method before the actual classifying process to deal with imbalanced data and compare them to SVM with SMOTE and normal SVM. The highest accuracy using a combination between SVM and KNN- Undersampling reached 63.83% which give slightly better result compared to a combination between SVM and SMOTE or normal SVM.