Deep Link Strength Prediction: Leveraging line graph transformations and neural networks

Abstract

Predicting link strengths in complex networks is a fundamental challenge, crucial for understanding network dynamics and optimizing real-world applications. Traditional approaches often rely on shallow structural features, limiting their ability to model intricate dependencies. To address these limitations, we propose Deep Link Strength Prediction (DLSP), a novel framework that integrates line graph transformations with graph convolutional networks (GCNs) to enhance the predictive capability of link weight estimation. DLSP redefines the task by transforming edge-centric information into node-level representations, facilitating effective learning of complex structural patterns. DLSP follows a multi-phase approach: first, a localized subgraph around the target link is extracted and encoded using a weighted node labeling scheme, preserving local structural and attribute-driven properties. Next, the labeled subgraph undergoes a line graph transformation, mapping link dependencies into node representations, thereby enabling a structured embedding space. A GCN is then employed to extract rich hierarchical representations, capturing both micro and macro-level graph structures. Finally, these learned embeddings are passed through a dense neural network to estimate the target link strength, framing the problem as a continuous-valued regression task. Unlike existing methods that rely on handcrafted features or isolated node embeddings, DLSP explicitly models link dependencies through graph-aware transformations, leading to superior predictive performance. Extensive experiments conducted on six diverse network datasets demonstrate that DLSP consistently outperforms state-of-the-art methods, showcasing its robustness, scalability, and potential for real-world applications.