A Topical Review of Graph Embedding in Graph Neural Networks

Graph embeddings map graph-structured data into vector spaces for machine learning tasks. In Graph Neural Networks (GNNs), these embeddings are computed through message passing and support tasks such as node classification, link prediction and community detection across several application domains. Prior reviews and benchmarking studies often focus on accuracy or scalability alone and do not examine structural preservation or the effect of model design on network distortion. This limits comparability and leaves open questions on how embeddings reflect graph topology in GNNs models. This topical review examines how graph embeddings are generated within GNN architectures and how model design choices affect both classification performance and graph structural preservation. The present work introduces a benchmarking framework that evaluates GNNs across different dimensions, including local and global structure preservation, complex relationships and scalability. The framework uses distortion metrics, automated parameter search and sensitivity analysis to provide a more complete view of model behavior. A controlled set of experiments was conducted using shared datasets and common hyperparameter spaces. The results show that GAT maintains stable performance across configurations, whereas GCN and GraphSAGE are more sensitive to parameter choices. The main takeaway is that reliable benchmarking of GNN-based embeddings requires a multidimensional analysis. Model behavior varies across datasets, and no architecture performs better across all conditions.