Merit: multi-level graph embedding refinement framework for large-scale graph

Abstract The development of the Internet and big data has led to emergence graphs as an important representation structure in various real-world scenarios. However, size increases, computational complexity memory requirements pose significant challenges for graph embedding. To address this challenge, paper proposes a multilevel embedding refinement framework (MERIT) based on large-scale graphs, using spectral distance-constrained coarsening algorithms improved convolutional neural network model that addresses over-smoothing problem by incorporating initial values identity mapping. Experimental results datasets demonstrate effectiveness MERIT, with average AUROC score 8% higher than other baseline methods. Moreover, node classification task 126,825 nodes 22,412,658 edges, improves quality while enhancing runtime 25 times. experimental findings highlight superior efficiency accuracy proposed approach compared