Detection of Cyberbullying in Social Media Streams: Metaheuristic-Tuned Temporal Fusion Network for High- Accuracy Text Interpretation
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Online communication growth has led to an increase in cyberbullying incidents while existing systems find it difficult to interpret evolving and abusive communication, creating the need for a better and context, and time-sensitive detection system. The current research is to introduce metaheuristic - tuned Temporal Fusion Network supported by specialized module designed as Dynamic Vortex Search - driven temporal attention with Long Short-Term Memory (DVS-TA-LSTM) to provide the high accuracy text interpretation for Cyberbullying detection. The dataset to identify cyberbullying was obtained from different types of social media (SM) platforms such as YouTube, Wikipedia Talk pages, Twitter, and Kaggle. Pre-processing consists of token normalization, removal of contextual noise, and textual data lexical correction. Feature extraction uses a Bidirectional Encoder Representations from Transformers (BERT) embedding model to represent part of the essential semantic meaning, contextual relationships, and information of sentiment in the text. The proposed framework integrates these components into a unified flow where cleaned data is embedded, processed through the fusion network, and further refined by the DVS-TA-LSTM module. Within this module, the DVS optimizer dynamically adapts LSTM hyper-parameters and efficiently detects cyberbullying by capturing sequential dependencies, and temporal attention reveals significant time- dependent patterns. The overall architecture is tailored to elevate cyberbullying detection accuracy across streaming environments. When compared to standard DL classifiers, the classifiers had the highest accuracy of 97.28%, the highest precision of 93.45%, the highest recall of 95.78%, the highest F1-measure of 94.55%, and the highest specificity of 97.54%, according to the Python simulation tool for experimental assessment. The framework concludes with the capability to support real-time, scalable, and context-aware cyberbullying mitigation.
Fields of Study & Subjects
Hate Speech and Cyberbullying Detection
Sentiment Analysis and Opinion Mining
Spam and Phishing Detection
Publication details
- Journal / source
- International Research Journal of Multidisciplinary Technovation
- Pages
- 251–266
- Publication date
- 28 March 2026
- Language
- EN