基于粗糙图的图卷积神经网络算法  

Graph Convolutional Neural Network Algorithm Based on Rough Graphs

作　　者：潘柏儒 丁卫平[1] 鞠恒荣 黄嘉爽 程纯[1] 沈鑫杰 耿宇 PAN Bairu;DING Weiping;JU Hengrong;HUANG Jiashuang;CHENG Chun;SHEN Xinjie;GENG Yu(School of Information Science and Technology,Nantong University,Nantong 226019)

机构地区：[1]南通大学信息科学技术学院,南通226019

出　　处：《模式识别与人工智能》2022年第9期827-838,共12页Pattern Recognition and Artificial Intelligence

基　　金：国家自然科学基金项目(No.61976120,62006128,62102199);江苏省自然科学基金项目(No.BK20191445);江苏省双创博士计划(No.(2020)30986);江苏省高等学校自然科学研究重大项目(No.21KJA510004);江苏省高等学校自然科学研究面上项目(No.20KJ B520009);南通市科技局基础科学研究项目(No.JC2020141,JC2021122);中国博士后科学基金项目(No.2022M711716);教育部人文社会科学研究青年基金项目(No.21YJCZH013)资助。

摘　　要：图卷积神经网络在解决节点分类问题时,使用拓扑图刻画节点间关系,并根据该拓扑图进行节点特征更新.然而,传统的拓扑图只能刻画节点之间的确定关系(即连接边权重为固定值),忽略真实世界中广泛存在的不确定性.这些不确定性不仅影响节点之间的关系,同时影响模型最终的分类性能.为了克服该缺陷,文中提出基于粗糙图的图卷积神经网络算法.首先,使用上下近似理论和传统拓扑图的边理论构造粗糙边,在粗糙边中使用成对出现的最大-最小关系值刻画节点之间的不确定关系,从而构建粗糙图.然后,设计基于粗糙图的可端到端训练的神经网络架构,将使用粗糙权重系数训练后的粗糙图输入图卷积神经网络,使用这些不确定信息更新节点特征.最后,根据这些学习的节点特征进行节点分类.在真实数据上的实验表明,文中算法可提高节点分类的准确率.A topology graph is utilized to portray the relationship between nodes and node features are updated on the basis of the topology graph,when the graph convolutional neural network is applied for solving node classification problems.However,traditional topology graph can only portray definite relationships between nodes,i.e.fixed values of connected edge weights,ignoring the widespread uncertainty in the real world.The uncertainty affects not only the relationship between nodes,but also the final classification performance of the model.To overcome this defect,a graph convolutional neural network algorithm based on rough graphs is proposed.Firstly,the rough edges are constructed via the upper-lower approximation theory and the edge theory of traditional topology graph,and the uncertain relationships between nodes are inscribed in rough edges using maximum-minimum relationship values coming in pairs to construct a rough graph.Then,an end-to-end trainable neural network architecture based on rough graphs is designed,the rough graphs trained with rough weight coefficients are fed into the graph convolutional neural network,and the node features are updated based on the uncertain information.Finally,nodes are classified according to the learned node features.The experiments on real dataset show that the proposed algorithm improves the accuracy of node classification.

关 键 词：图卷积神经网络 拓扑图 粗糙集 粗糙图 不确定关系 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]