Improvement in Cyberbullying Detection Using Deep Learning

Abstract

In this digital world, avoiding the internet completely is not possible. Due to anonymity gained in using the internet, trolls and abuses for fun or for sadist reasons have increased that in turn affects the victim’s health in a negative way. The amount of data generated is huge, thus an automated system to detect cyberbullying and act accordingly is required. The recently proposed techniques may generate certain bias for a particular group of people. This can be due to the distribution of samples in training data, based on attribute such as ethnicity, language, religion, etc. In this work, a fusion of Fairness Constraint (FC) and class balancing techniques is proposed. The proposed fusion aims to handle both bias problem and class imbalance problem at the same time. The performance of classweight, undersampling (Nearmiss, Edited Nearest Neighbor, One Sided Selection and Neighborhood Cleaning Rule) and oversampling (SMOTE, Borderline-SMOTE, SVMSMOTE and ADASYN) techniques are observed for three datasets. The performance of the fusion is compared to baseline model, FC model, and model with class balancing technique.