Variational Graph Convolutional Networks for Dynamic Graph Representation Learning

The ubiquitous and ever-evolving nature of cyber threats demands innovative approaches that can adapt to the dynamic relationships and structures within network data. Traditional models struggle to adapt to the constantly changing nature of network traffic, where both structural dependencies and temporal evolution must be accurately captured to detect anomalies and predict future threats. To address the challenges, this research introduces V-GCN (Variational Graph Convolutional Network), a new model that integrates the probabilistic latent space modelling of Variational Autoencoders (VAEs) with the structural learning capabilities of Graph Convolutional Networks (GCNs). The proposed model is designed to capture both temporal dependencies and uncertainties inherent in dynamic networks, and as such, it is highly suitable for tasks such as link prediction and node classification. The proposed hybrid model encodes node features into a probabilistic latent space using a VAE encoder and refine the representations using GCN layers, that aggregates structural information from neighbouring nodes. The integration of variational inference with graph convolution enables V-GCN to adapt to the dynamic evolution of network traffic and measure the uncertainties in node and edge relationships. The DynKDD dataset, a dynamic adaptation of the NSL-KDD dataset, is developed in this research to evaluate the model performance. The dataset introduces temporal dynamics into the conventional NSL-KDD dataset, enabling the application of advanced graph-based learning models such as V-GCN. Experimental evaluation indicates that V-GCN significantly outperforms baseline models such as GCNs, Graph Sample and Aggregation (GraphSAGE), and Graph Attention Networks (GATs). In node classification, V-GCN achieved a 10% higher F1-score (0.845), with precision reaching 83.7%, and a balanced accuracy of 84.2%, underscoring its ability to handle uncertainty and adapt to changing network structures in dynamic environments. V-GCN achieved a 15% improvement in AUC-ROC (0.98), a 12% increase in average precision (0.9357), and a 14% higher F1-score (0.8196) in link prediction tasks compared to baseline models. The V-GCN's integration of probabilistic modelling and graph convolution sets a new benchmark for dynamic network traffic analysis, providing a superior solution to real-world challenges in cybersecurity, social network analysis and beyond.