FastAlign Boosts Scalability for Network Alignment Algorithms

Network alignment, a critical task in data science for identifying node correspondences across different networks, often suffers from scalability issues. Existing high-accuracy methods typically rely on computationally expensive operations involving dense matrices. FastAlign addresses this by introducing a scalable, sparsity-aware framework that preserves the original optimal transport (OT) formulation. Instead of altering the alignment model, FastAlign reinterprets its computations as a series of mixed sparse-dense operations. The framework combines sparsity-aware graph computation with specialized kernel fusion, including a custom SpMM (Sparse Matrix-Matrix multiplication) kernel. This approach allows FastAlign to achieve alignment quality comparable to state-of-the-art OT-based methods while delivering substantial runtime reductions, up to 9.45x on CPU and 32.54x on GPU.