节点属性增强的图自编码器

Graph auto-encoder with node attribute enhanced

针对传统图自编码器的解码方法忽略节点属性作用的问题,提出一种联合重建图结构和属性信息的节点属性增强的图自编码器(NEGAE)模型。模型在编码器部分,采用图卷积神经网络进行图节点数据的特征提取,获得其节点表示;在解码器部分,一方面采用内积方式对图结构进行重建,另一方面采用反卷积的方式对节点属性进行重建;最后,将结构信息和节点属性信息的重建误差融合到一个统一的损失函数中进行优化。在Cora、Citeseer、Pubmed数据集上的结果表明:该模型在链路预测任务中的ROC曲线下面积(AUC)分别达到91.19%、90.27%、96.69%;聚类任务中的聚类准确度(ACC)分别达到60.31%、50.60%、66.79%,说明NEGAE方法在各种学习任务上均取得了良好的性能。

Aiming at the problem that the decoding methods of traditional graph auto-encoder ignores the effect of node attribute,a graph auto-encoder(NEGAE)model with node attribute enhanced which contains both structure reconstruction and node attribute reconstruction is proposed.In the encoder part,the model uses graph convolutional neural network to extract features of nodes,and obtains its representations.In the decoder part,on the one hand,the inner product method is used to reconstruct graph structure,on the other hand,the deconvolution method is used to reconstruct the node attributes.Finally,the reconstruction error of structure information and node attribute information are merged into a unified loss function framework for optimization.The results on Cora,Citeseer and Pubmed datasets show that the area under Roc curve(AUC)of the model on link prediction task reaches 91.19%,90.27%and 96.69%,respectively;the clustering accuracy(ACC)on clustering task reaches 60.31%,50.60%and 66.79%,respectively.It illustrates that NEGAE method has achieved good performance on various learning tasks.