A Comparative Analysis of Machine Learning Techniques for Cyberbullying Detection on Twitter

Analysis

Why This Paper Matters

Cyberbullying is a pervasive issue on social media, and automated detection is crucial for intervention without relying on victim reports. This paper addresses a practical gap by compiling a large, global dataset and systematically comparing multiple machine learning classifiers. The findings are significant because they demonstrate that a relatively simple model like Logistic Regression can outperform more complex ensemble methods, which has implications for computational efficiency and real-world deployment.

The study's focus on a global dataset (37,373 tweets) is a strength, as it increases the diversity of language and context, making the results more generalizable than studies using single-region data. By evaluating seven classifiers under the same conditions, the paper provides a fair comparison that is often missing in literature, offering a clear benchmark for future research.

Technical Contributions

The paper's main technical contributions include:

• Dataset creation: A new global dataset of 37,373 unique tweets, which is a valuable resource for the research community.
• Comprehensive evaluation: Systematic comparison of seven classifiers (LR, LGBM, SGD, RF, ADB, NB, SVM) using four standard metrics.
• Performance insights: Identification of trade-offs between precision and recall across classifiers, with SGD achieving highest precision and SVM achieving perfect recall.
• Practical recommendation: Demonstrating that Logistic Regression offers the best balance of accuracy and F1 score, suggesting that simpler models can be effective for this task.

Results

The experimental results show that Logistic Regression achieved a median accuracy of 90.57%, the highest among all classifiers. It also achieved the best F1 score of 0.928, indicating a strong balance between precision and recall. SGD achieved the highest precision at 0.968, meaning it had the fewest false positives, while SVM achieved a perfect recall of 1.00, meaning it identified all actual cyberbullying instances but likely at the cost of precision. These results highlight the importance of selecting metrics based on the application's needs—e.g., high recall for safety, high precision for reducing false alarms.

Significance

This paper contributes to the field of social media content moderation by providing a robust comparison of machine learning techniques for cyberbullying detection. The finding that Logistic Regression performs competitively with more complex models is practically significant, as it suggests that lightweight models can be deployed in real-time systems with lower computational costs. The dataset and methodology also serve as a baseline for future research, encouraging further exploration into feature engineering and deep learning approaches. Overall, this work supports the development of automated tools to mitigate cyberbullying, which has positive societal implications.