Detection of online hate through fuzzy-based optimised machine learning models and data augmentation

Abstract

The proliferation of cyber-hate alongside the transformation of global communication via social media platforms has prompted researchers to explore effective detection methods. This study critically evaluates the limitations of prevailing Machine Learning (ML) and Deep Learning (DL) techniques, particularly in handling sentiment-oriented data inherent in online hate speech. Traditional classification methods such as Naive Bayes, Logistic Regression, Convolutional Neural Networks, and Recurrent Neural Networks may inadequately capture the nuanced nature of this phenomenon, necessitating a more critical thinking perspective for precise classification.

While DL methods have shown remarkable success in various natural language processing tasks, including hate speech detection, they often come with significant computational costs and training time. This study acknowledges the imperative for efficiency in real-world applications, where computational resources may be limited, and timely responses to cyber-hate are crucial.

In response, this research conducts a comprehensive analysis utilizing two machine learning classifiers, namely Multinomial Naive Bayes (MNB) and Logistic Regression (LR), across four distinct datasets. Methodologies are meticulously selected to prioritize interpretability, efficiency, and effectiveness in handling sentiment-oriented data. Additionally, Particle Swarm Optimization and Genetic Algorithms are integrated to optimize classifier performance by fine-tuning parameters, leveraging their capacity to efficiently explore solution spaces and enhance classification accuracy. Fuzzy Logic is employed to augment classifier robustness against noisy data by encapsulating inherent uncertainty and ambiguity in textual content.

Acknowledging concerns regarding competition with deep learning models and the imperative for efficiency, the chosen approaches prioritize interpretability without compromising performance. Leveraging bio-inspired optimization techniques, the aim is to achieve competitive performance while reducing computational burden and training time compared to deep learning approaches. This strategic approach not only enhances cyber-hate detection capabilities but also addresses practical challenges associated with deploying sophisticated models in real-world scenarios. Moreover, the study explores the application of Generative Adversarial Networks (GAN) as a potential solution to address highly imbalanced datasets. Comparative analysis reveals that GANs offer unique advantages in effectively detecting hate-related data while mitigating risks of overfitting. By presenting concrete results and insights from this comparison, the author underscores the specific contributions of GANs in advancing cyber-hate detection.

Overall the aim of this research is to enhance the detection of online-hate through the integration of machine learning , bio-inspired optimization techniques, fuzzy logic, and generative adversarial networks. The research specifically seeks to address the limitations of existing classification methods and deep learning techniques in accurately identifying and categorizing online hate.

To achieve this, the study proposes several hypotheses. The first hypothesis suggests that using MNB and LR classifiers, optimized using PSO and GA, will lead to improvements in the accuracy and interpretability of cyber-hate detection compared to traditional machine learning and deep learning approaches. The second hypothesis suggests that incorporating fuzzy logic into the classification process will enhance the robustness of classifiers against noisy data, thereby improving overall performance in detecting online hate. Lastly, the third hypothesis states that the application of GANs will effectively address imbalanced datasets in cyber-hate detection, offering unique advantages over conventional methods and mitigating the risks of overfitting. These hypotheses collectively aim to advance the efficacy of cyber-hate detection systems, making them more reliable and efficient in a real-world context.

Date of Award	23 May 2024
Original language	English
Awarding Institution	Northumbria University
Supervisor	Biju Issac (Supervisor)