Network backbone extraction is essential for simplifying complex systems while preserving critical structural features. While traditional backbone extraction methods rely on statistical tests and topological attributes, this research presents a novel approach using embedding-based link prediction techniques. We conduct a comparative analysis with classical methods across several real-world networks to assess the effectiveness of different embedding methods. Our findings show that Graph Autoencoders (GAE) and Laplacian Eigenmaps (LEM) based-backbone methods effectively capture regional connections in a hub-and-spoke air transportation network. In contrast, the High Salience Skeleton method excels in preserving overall network connectivity and reachability. Among the methods analyzed, DeepWalk preserves the weight distribution best, and LEM is most effective in maintaining the degree distribution. These results suggest that selecting a backbone extraction method should depend on the analysis’s specific network characteristics and objectives. This approach offers valuable insights for transportation, social, and other types of networks.