Anomaly Detection in Social Networks Using Deep Learning and Attention Graphs
Anomaly Detection in Social Networks Using Deep Learning and Attention Graphs
Zahra rohani1
1) Instructor, Computer Department, Shirvan Branch, Islamic Azad University, Shirvan, Iran
IOI Code
:
AAHD-ABFHEF
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
Social networks have become an integral part of everyday life today, but these same platforms are facing serious challenges from anomalous and destructive activities such as social bots, fake news, and public opinion manipulation. This paper provides a comprehensive review of anomaly detection methods in social networks with a special emphasis on approaches based on deep learning and attention graphs. Given the dynamic, heterogeneous, and large-scale nature of social networks, traditional methods based on hand-crafted features and classical machine learning do not have the necessary ability to deal with complex and evolving attack patterns. In this study, first, frameworks based on graph neural networks (GNN) including graph convolutional networks (GCN) and graph attention networks (GAT) are reviewed. Then, advanced architectures based on attention mechanisms and transformers that are capable of modeling long-range dependencies in graph structures are analyzed. In particular, novel methods such as ST-MVAN (Spatio-Temporal Multi-View Attention Network) and BotLGT (Robot Detection Based on Graph Transformers and Large Languages) are investigated in depth. This study shows that hybrid approaches that combine large language models (LLM) with attention graphs are able to achieve accuracy of up to 98.5% in social robot detection. Also, existing challenges such as scalability, interpretability, and data imbalance are reviewed and future research directions are outlined. The findings of this paper show that the use of sparse attention mechanisms, multi-exponential adaptive fusion, and error reconstruction-based autoencoders provides a significant improvement in the AUC index of up to 12.26% compared to the baseline methods.
Keywords : anomaly detection social networks deep learning attention graph graph neural network robot detection fake news natural language processing