Abstract Complex networks are fundamental to understanding systems across diverse domains. However, their increasing size and complexity pos
Abstract Complex networks are fundamental to understanding systems across diverse domains. However, their increasing size and complexity pose challenges for analysis and scalability. Backbone extraction simplifies networks by preserving key structures while reducing complexity. Traditional methods include statistical and structural approaches, with statistical methods often lacking interpretability and structural methods offering limited flexibility. This paper introduces a novel framework for backbone extraction using similarity-based link prediction methods in weighted and unweighted networks. It aligns Low Similarity backbones (LS-backbones) with traditional statistical methods and High Similarity backbones (HS-backbones) with structural methods. LS-backbones retain edges with low similarity scores, emphasizing weak or unexpected connections. HS-backbones preserve edges with high similarity scores, focusing on strong and cohesive structures. In this study we employ three similarity functions: Preferential Attachment (local), Local Path Index (quasi-local), and Shortest Path Index (global). Quantitative evaluation across 18 networks shows that: LS-backbones preserve more nodes and highlight peripheral structures but struggle with fragmentation at lower edge retention levels. HS-backbones maintain superior connectivity and interconnectivity, making them ideal for robust and cohesive networks. A qualitative evaluation using the global air transportation network highlights the framework’s versatility. While the LS-methods aligns with statistical methods they are more interpretable due to their clear and well-defined similarity assumptions. Similarly, although the High Salience Skeleton is interpretable, the framework demonstrates greater flexibility by varying the similarity function to capture diverse network complexities. For instance, the Low Similarity Preferential Attachment backbone uncovers weak peripheral links, enhancing regional connectivity. Low Similarity Shortest Path backbone identifies critical connections vital for maintaining isolated regions, while Low Similarity Local Path backbone highlights non-redundant connections in constrained areas. High Similarity Preferential Attachment backbone emphasizes robust global links between major hubs, High Similarity Shortest Path backbone ensures efficient direct paths, and High Similarity Local Path backbone reveals redundant connections that strengthen localized cohesion. This study demonstrates the versatility of similarity-based link prediction for extracting meaningful backbones. Future research will explore the synergy between LS-backbones and HS-backbones, alongside advanced link prediction methods, to further expand the framework’s applicability.
