An ensemble deep learning model based on graph neural networks for intelligent caching in next generation data centers

Abstract Repeated user requests to edge servers increase traffic load and content-delivery latency. Although recommender-driven caching is promising, it still faces challenges of scalability, sparsity, cold-start, and limited generalization. We subsequently develop an error-aware ensemble method that adaptively weights predictors to strike a trade-off between individual-model robustness and domain-specific bias mitigation across varied datasets. By taking advantage of the generalization capability in Graph Neural Network (GNN), we propose a hybrid content-selection policy for edge cache placement. First, Light Graph Convolutional Network (LightGCN) and Knowledge Graph Attention Network (KGAT) are employed to jointly capture high-level semantic relationships and fine-grained user–item interaction patterns. Second, Mean Squared Error Inverse (MSEI)-based weighting scheme is introduced to proportionally combine LightGCN and KGAT, thereby balancing their complementary strengths and enhancing overall model performance. Compared with the baseline methods PMF and SVD, the proposed model yields significant gains of 13.6% and 10.16% in Precision@10, and 13.84% and 8.82% in NDCG@10 on the MovieLens dataset, highlighting its effectiveness in achieving adaptive and efficient caching placement for next-generation edge data centers.