Temporal Dual-Attention Graph Network for Popularity Prediction in MEC Networks

Mobile Edge Caching (MEC) emerges as a key technology in communication networks to minimize latency in handling user requests for high-demand, data-intensive services. Given the storage constraints of edge servers, it becomes imperative to dynamically cache the most popular content in a highly optimized and seamless fashion. To address this critical challenge, we leverage the capabilities of Graph Neural Networks (GNNs), renowned for their proficiency in modeling intricate interdependencies among users and content, inter-user interactions, and inter-content correlations. Existing GNN-based cashing solutions failed to properly manage computational complexity of the structural learning process, but instead mainly focused on enhancing the temporal learning aspects. Since edge servers are mainly resource-limited, it is crucial to reduce the complexity of models to reduce the train and test time. The paper addresses this gap. More specifically, capitalizing on the adaptive and scalable characteristics of GNNs, we propose the Temporal Dual Graph Attention Network (TDA-GNN) for efficient content popularity prediction in MEC networks. The proposed TDA-GNN framework employs two distinct attention units, recognizing the varied significance of neighbouring nodes and edge features in shaping node embeddings. Comparative analysis against a baseline architecture reveals that the proposed TDA-GNN framework boosts cache-hit rate. These findings underscore the potential of dual-attention mechanisms to significantly enhance MEC performance, promising smoother, faster access to content for mobile users.