Hybrid ensemble approaches to online harassment detection in highly imbalanced data

Abstract

Online harassment is a major threat to users of social media platforms, especially young adults and women. It can cause mental illnesses and impacts deeply and negatively economic institutions experiencing cyberbully attacks by losing their credibility and business. This makes automatic detection of online harassment extremely important. Most of current studies within this context apply machine-learning algorithms that assume balanced class distribution. However, this assumption does not hold for most real datasets. This research provides a comprehensive investigation of various approaches that combine diverse techniques under three dimensions: feature representation, imbalanced data handling, and supervised learning. For the first dimension, three word-embedding models have been considered, namely: word2vec, Glove, and SSWE. For the other two dimensions, nine techniques for balancing skewed class distributions have been employed to feed several learning models. In particular, resampling methods, cost-sensitive learning, and Weight-Selection strategy-based methods have been used with deep neural networks. The ultimate goal of this study is to evaluate the potential of using such hybrid approaches to handle the online harassment detection task efficiently using highly-imbalanced Twitter data and to select the best combination concerning the intended purpose. An extensive comparative study has been conducted, and the results have been discussed in terms of three evaluation metrics widely used for imbalanced classification. As main findings, Glove has been found as the best feature representation and some combinations as the best performing most notably LSTM and BLSTM with cost-sensitive learning and VL strategy.