Improving detection accuracy of politically motivated cyber-hate using heterogeneous stacked ensemble (HSE) approach

Abstract

The surge in cyber-hate crimes is largely fuelled by the popularization of social media platforms. On that note, cyber-hate has become an increasing concern for most countries, especially those that are practising democracy. Studies on the influence of social media (SM) on political discourse have now become an important research area due to the rising trends of SM politics. It becomes necessary to address this problem using automated social intelligence. To tackle this concern, the researchers built a novel heterogeneous stacked ensemble (HSE) classifier for detecting politically motivated cyber-hate on Twitter. We constructed a heterogeneous stacked ensemble with eight baseline estimators. In the proposed methodology, the researchers employed TF-IDF for feature vectorisation. The researchers used Twitter API for data scraping to harvest tweets during a gubernatorial election in Nigeria for the training and evaluation of the stacked ensemble model. A total of 15,502 tweets were collected and after some preliminary cleaning, 5876 tweets were manually labelled as hate (1) or non-hate (0). The coded tweets contain 16.87% hate and 83.13% non-hate tweets. This article has three contributions – a critical review of literature on the detection of politically motivated cyber-hate, the building of a new dataset and the proposed stacked ensemble method. Two other public datasets (Kaggle and HASOC) were used to test the performance of our method. The F1-score metric was employed for comparison. Our method is better by 12% on the Kaggle and 4% on the HASOC datasets. We are working on more data for deep learning experiments.