Mg-SubAgg: Multi-granularity Subgraph Aggregation with topology for GNN

作     者：Zhang, Xiaoxia Ye, Mengsheng Zhang, Yun Liu, Qun Wang, Guoyin Wu, Kesheng 

作者机构：Chongqing Univ Posts & Telecommun Chongqing Key Lab Computat Intelligence Chongqing 400065 Peoples R China Chongqing Univ Posts & Telecommun Key Lab Big Data Intelligent Comp Chongqing 400065 Peoples R China Chongqing Univ Posts & Telecommun Key Lab Cyberspace Big Data Intelligent Secur Minist Educ Chongqing 400065 Peoples R China Lawrence Berkeley Natl Lab Sci Data Div Berkeley CA 94720 USA 

出 版 物：《INFORMATION SCIENCES》 (信息科学)

年 卷 期：2024年第677卷

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 08[工学] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：National Natural Science Foundation of China [61936001, 62221005, 62136002, 62233018, 12201089] Natural Science Foundation of Chongqing [cstc2021 ycjh-bgzxm0013, cstc2019jcyj-cxttX0002, cstb2022nscq-msx0226, cstc2022ycjh-bgzxm0004, CSTB2023NSCQ-LZX0006] Key Cooperation Project of Chongqing Municipal Educa-tion Commission [HZ2021008] Science and Technology Research Program of Chongqing Education Commission of China [KJQN202200513] Innovation Projects for Studying Abroad and Returning to China of Chongqing Municipal Bureau of Human Resources and Social Security of China [cx2023097] 

主　　题：Graph Neural Network (GNN) Explainability of GNN Subgraph Aggregation Improved Shapley value Node importance 

摘      要：Although graph neural networks (GNNs) work well on graph data, they are black-box models that lack of reliable explanations for their predictions. We propose a multi-granularity subgraph aggregation method based on graph topology to explain GNNs. Specifically, given a trained GNN model and an input graph, our method constructs a subgraph by heuristics from fine-grained to coarse-grained and sorts the subgraph nodes to obtain subgraph and node-level explain. Furthermore, we propose an improved Shapley value as a heuristic function for the search algorithm, which strikes a balance between the time complexity and accuracy. Finally, experimental results on both synthetic and real datasets demonstrate that our method achieves best performance on seven datasets, quantifying the influence of individual nodes on prediction results and providing more reliable explanations.