Comparative Analysis of Various Machine Learning Algorithms to Detect Cyberbullying on Twitter Dataset

Abstract

The advent of the digital era has seen the rise of social media as an alternative mode of communication. The usage of social media platforms for facilitating contact between individuals has become widespread. As a direct result of this, conventional modes of communication have been replaced by digital modes, thanks to social media. Because of the growing prevalence of cyberbullying, this digital development on social media platforms is a significant problem that must be addressed. Bullies have various options to harass and threaten individuals in their communities because of the platforms that are already available. It has been argued that a number of different tactics and approaches may be employed to combat cyberbullying via the use of early identification and alerts to locate and/or protect victims of cyberbullying. Methods from the field of machine learning (ML) have seen widespread use in the search for language patterns used by bullies to cause damage to their victims. This research paper analyzes standard supervised learning and ensemble machine learning algorithms. The ensemble technique utilizes random forest (RF) and AdaBoost classifiers, whereas the supervised method uses Gaussian Naive Bayes (GNV), logistic regression (LR), and decision tree (DT). We use the dataset to train and evaluate our binary classification model to classify abusive language as bullying or non-bullying and extract Twitter features using term frequency-inverse document frequency (TF-IDF). Downloaded the dataset from Kaggle. This paper analyzes each machine learning algorithm. Ensemble-supervised algorithms outperformed standard supervised algorithms in the analysis. With a dataset, the random forest classifier performed best with 92% accuracy, while the Naive Bayes classifier performed worst with 62% accuracy.