Score-based progressive edge mask graph autoencoder

Graph neural networks (GNNs) have demonstrated exceptional capabilities in graph representation learning, particularly for link prediction tasks essential in various domains such as social network analysis, recommendation systems, and biological network research. This paper addresses critical challenges in link prediction, including the inefficiencies of random masking strategies, fixed masking ratios, and limited structural representation in existing methods. We propose a novel framework, Score-based Progressive Edge Masking Graph Autoencoder (SPEM-GAE), which employs edge importance scores to guide a dynamic and progressive masking strategy. SPEM-GAE preserves crucial structural information in early training stages while gradually increasing the masking ratio to enhance robustness to less critical edges. A reparameterized sampling mechanism ensures end-to-end optimization, and a dual-decoder architecture captures both global structural patterns and local connectivity details by reconstructing the adjacency matrix and degree information. Extensive experiments on three benchmark citation datasets (Cora, CiteSeer, and PubMed) validate the superior performance of SPEM-GAE over state-of-the-art baselines, achieving improvements in AUC scores. Results show that SPEM-GAE effectively balances local and global graph structure learning, demonstrating robustness and adaptability. This work highlights the importance of progressive masking strategies guided by edge importance and opens avenues for further exploration in complex and dynamic graph structures.