AnoEdgePred: A Novel Method for Detecting Anomalous Edges in Social Networks
- Pages 205-221
The social network has grown exponentially, connecting people globally and evolving dynamically. Detecting anomalies within these networks is important as they may indicate malicious activity or error. Traditional algorithms like Isolation Forest and One-Class SVM focus on existing structures and overlook new connections, leading to inadequate handling of dynamic changes. Similarly, recent algorithms like Graph Convolutional Networks (GCNs) and Graph Autoencoders (GAEs), while effective for static graphs, may struggle with computational efficiencies. To address these issues, we introduce AnoEdgePred, a novel method that combines Graph neural networks (GNNs) with link prediction and network analysis techniques to identify anomalies in newly formed connections. The method predicts future edges and analyses their structure to identify anomalies such as stars, cliques, dominant edges, or bottleneck edges. AnoEdgePred provides a comprehensive analysis of edge features and structural deviations, often overlooked by existing methods. We evaluated AnoEdgePred by comparing its performance with four different algorithms across datasets, like Facebook, Enron, Slashdot, Polblogs, and synthetic powerlaw network, using different performance metrics. Results show that AnoEdgePred significantly improves detection accuracy by 5−60% and achieves substantial gain in other metrics like precision, recall, F1-score, and AUC, making it a suitable solution for evolving social networks.
Taylor & Francis Online Top Full Article Figures & data References Supplemental Citations Metrics Reprints & Permissions View PDF(open in a new window) Share Formulae display:MathJax Logo? Abstract The social network has grown exponentially, connecting people globally and evolving dynamically. Detecting anomalies within these networks is important as they may indicate malicious activity or error. Traditional algorithms like Isolation Forest and One-Class SVM focus on existing structures and overlook new connections, leading to inadequate handling of dynamic changes. Similarly, recent algorithms like Graph Convolutional Networks (GCNs) and Graph Autoencoders (GAEs), while effective for static graphs, may struggle with computational efficiencies. To address these issues, we introduce AnoEdgePred, a novel method that combines Graph neural networks (GNNs) with link prediction and network analysis techniques to identify anomalies in newly formed connections. The method predicts future edges and analyses their structure to identify anomalies such as stars, cliques, dominant edges, or bottleneck edges. AnoEdgePred provides a comprehensive analysis of edge features and structural deviations, often overlooked by existing methods. We evaluated AnoEdgePred by comparing its performance with four different algorithms across datasets, like Facebook, Enron, Slashdot, Polblogs, and synthetic powerlaw network, using different performance metrics. Results show that AnoEdgePred significantly improves detection accuracy by 5−60% and achieves substantial gain in other metrics like precision, recall, F1-score, and AUC, making it a suitable solution for evolving social networks
Edge prediction Graph Neural Networks (GNNs) Link prediction Network analysis techniques Social network analysis