Enhancing the Efficiency of Unsupervised Network Alignment Using Quotient Graph

Network alignment, a foundational technique for cross-domain applications such as recommendation systems and knowledge fusion, faces significant challenges in balancing computational efficiency and alignment accuracy. Although graph neural networks (GNNs) effectively capture structural and semantic relationships, their computational intensity—stemming from multi-layer matrix operations and high memory consumption—severely limits scalability. To address these limitations, this paper proposes ENAMOR (Efficient Network AlignMent via Quotient gRaph), an unsupervised framework incorporating three key components: 1) multi-scale representation learning that hierarchically aggregates local and global structural patterns through GNN layers; 2) embedding-driven graph coarsening via hashing-based quotient graph construction, reducing computational complexity by 60–80% while preserving topological and attribute information; and 3) Matched Neighborhood Consistency (MNC) optimization, which iteratively refines alignment matrices by enforcing structural congruence constraints. Extensive experiments on three real-world datasets (Douban Online-Offline, Allmovie-Imdb, ACM-DBLP) demonstrate that ENAMOR achieves a 13.47–94.56% reduction in runtime compared to state-of-the-art methods, alongside improvements of 0.27–13.47 percentage points in mean average precision (MAP) and 0.15–8.17 percentage points in precision@5. The framework effectively balances efficiency and accuracy, providing a scalable solution for cross-network analysis tasks.