文献类型：期刊文献

英文题名：Geometric and topological structure-induced large-scale graph learning for social and information networks

作者：Liu, Ganghao[1,2];Xiao, Ting[1,2];Wang, Zhe[1,2];Wang, Hanzheng[1,2]

机构：[1]Minist Educ, Key Lab Smart Mfg Energy Chem Proc, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Dept Comp Sci & Engn, Shanghai 200237, Peoples R China

年份：2026

卷号：173

外文期刊名：PATTERN RECOGNITION

收录：;EI(收录号：20255219788800);WOS:【SCI-EXPANDED(收录号:WOS:001651644800001)】；

基金：This work is supported by the Natural Science Foundation of China under Grant No. 62476087, the Shanghai Municipal Education Commission's Initiative on Artificial Intelligence-Driven Reform of Scientific Research Paradigms and Empowerment of Discipline Leapfrogging, the National Key Research and Development Program of China under Grant No. 2022YFB3203500, the Special Fund for Basic Scientific Research Operations in Central Universities, and the Natural Science Foundation of Shanghai under Grant No. 24ZR1416800.

语种：英文

外文关键词：Graph representation learning; Large-scale graphs; Node classification; Geometric and topologic structure

摘要：In recent years, remarkable progress has been made in graph representation learning, with continual refinements to model architectures improving their effectiveness. However, most existing approaches are primarily designed for small-scale graphs. As graph structures grow in size and complexity, conventional models struggle to capture spatial dependencies and often suffer from cyclic repetitive sampling during subgraph generation. To address these challenges, we propose a novel framework, termed the Geometric and Topologic Structure-induced Large-scale Graph Learning Method (GTSLGM), for representation learning on large-scale graphs. The primary goal of GTSLGM is to enhance spatial feature extraction and optimize subgraph sampling. Specifically, GTSLGM first reconstructs node neighborhoods based on geometric information to capture long-range structural dependencies. Then, a topology-guided random walk sampling strategy is introduced to mitigate cyclic repetitive sampling and ensure the generation of well-connected subgraphs. Finally, the sampled subgraphs are employed for downstream node classification tasks, leading to significant performance gains. Extensive experiments on four large-scale benchmark datasets demonstrate the superiority of GTSLGM over existing baselines in both accuracy and efficiency.