摘要
为了更好地学习网络中的高阶信息和异质信息,基于单纯复形提出单纯复形—异质图注意力神经网络方法—SC-HGANN。首先,用单纯复形提取网络高阶结构,将单纯复形转换为单纯复形矩阵;其次,使用注意力机制从特征单纯复形中得到异质节点的特征;再次,对同质和异质单纯复形矩阵进行卷积操作后,得到同质特征与异质特征,通过注意力算子进行特征融合;最后,得到目标节点的特征并将其输入到节点分类模块完成分类。与GCN、HGNN、HAN等基线方法相比,提出的方法在三个数据集上的macro-F1、micro-F1、precision和recall均有所提升。表明该方法能有效地学习网络中的高阶信息和异质信息,并能提升网络节点分类的准确率。
In order to better learn the high-order information and heterogeneous information in the network,this paper proposed a simplicial complex-heterogeneous graph attention neural network method based on simplicial complex(SC-HGANN).Firstly,it used simplicial complex to extract the high-order structure of the network,and took conversion from simplicial complex to simplicial complex matrix.Secondly,it applied the attention mechanism to obtain the feature of heterogeneous nodes from the features simplicial complex.Then,after convolution operation of homogeneous and heterogeneous simplicial complex matrix,homogeneous feature and heterogeneous feature took feature fusion to generate the feature of the target node by attention operator.Finally,the feature of the target node inputted the node classification module completes the classification.Compared with baseline methods such as GCN,HGNN and HAN,the SC-HGANN improves macro-F 1,micro-F 1,precision and recall on the three datasets.The results show that the SC-HGANN can effectively learn high-order information and heterogeneous information in the network,and improve the accuracy of node classification.
作者
陈东洋
郭进利
Chen Dongyang;Guo Jinli(Business School,University of Shanghai for Science&Technology,Shanghai 200093,China)
出处
《计算机应用研究》
CSCD
北大核心
2023年第4期1095-1100,1136,共7页
Application Research of Computers
关键词
单纯复形
高阶网络
注意力机制
图神经网络
节点分类
simplicial complex
higher-order network
attention mechanism
graph neural network
node classification