文献类型：期刊文献

英文题名：Graph neural network with feature enhancement of isolated marginal groups

作者：Wang, Jiahui[1];Guo, Yi[1,2,3];Wang, Zhihong[4];Wen, Xinxiu[1];Ni, Jianwei[1]

机构：[1]East China Univ Sci & Technol, Shanghai, Peoples R China;[2]Natl Engn Lab Big Data Distribut & Exchange Techn, Shanghai, Peoples R China;[3]Shanghai Engn Res Ctr Big Data & Internet Audienc, Shanghai, Peoples R China;[4]Minist Publ Secur, Res Inst 3, Shanghai, Peoples R China

年份：2022

卷号：52

期号：14

起止页码：16962

外文期刊名：APPLIED INTELLIGENCE

收录：;EI(收录号：20221411913883);WOS:【SCI-EXPANDED(收录号:WOS:000776951600002)】；

基金：This research is financially supported by the National Key Research and Development Program of China (grant number 2018YFC0807105), National Natural Science Foundation of China (grant number 61462073), and Science and Technology Committee of Shanghai Municipality (STCSM) (under grant numbers 17DZ1101003, 18,511,106,602, and 18DZ2252300).

语种：英文

外文关键词：Graph neural network; Graph data mining; Group partition; Label propagation; Isolated groups

摘要：Graph neural networks (GNN) have achieved remarkable success by combining feature and structure information. However, the over-smoothing phenomenon has always been a crucial issue in GNN models since the node representation will easily converge to the full graph representation with the increasing of convolutional layers. Our investigation also indicates that the isolated part significantly restricts the capability due to the meaningless structure and local over-smoothing problem. While current models treated all nodes as equal with the absence of effectiveness of structure at different locations. To facilitate this line of research, we innovatively propose the graph neural network with feature enhancement of isolated parts (GNN-FEIP) consisting of graph partition, graph construction, and strategic label propagation procedures. In GNN-FEIP architecture, all the nodes are partitioned into several groups according to their position and connectivity. The feature-level similarity graph is reconstructed for subsequent feature enhancement of isolated nodes. Afterward, the preliminary prediction of the original GNN model has been adjusted with the strategic label propagation, which both balances the feature and structure of the nodes at different positions in a comprehensive manner. The exhaustive experiments indicate that GNN-FEIP achieves impressive performance than other off-the-shelf models, especially in the case that the isolated part is of large proportion.