异质网中基于图卷积神经网络的链路预测方法

Link prediction approach based on graph convolution neural network in heterogeneous networks

综合考虑异质信息网络具有的复杂性和异质性的特点,提出一种异质网中基于图卷积神经网络(heterogeneous graph convolution neural network embedding,HeGCNE)的链路预测方法。针对经典图卷积神经网络逐层传递规则的不足,提出改进的逐层传递规则,对异质节点进行表征学习,融合对抗学习优化节点表征;在此基础上,利用节点的哈达玛积构造连边表征,将连边表征放入基于梯度提升树算法的二分类器,解决异质网络的链路预测问题。实验结果表明,改进后的方法可以有效提高链路预测的准确性和稳定性。

To address the problem of complexity and heterogeneity in heterogeneous information networks,a link prediction algorithm based on graph convolution neural network embedding in heterogeneous networks(HeGCNE)was proposed.To overcome shortcomings of the multi-layer rule of classical graph convolution neural network,an improved multi-layer rule was proposed to learn the embeddings of heterogeneous nodes.The quality of the node embeddings was improved through adversarial learning.Based on these,the Hadamard product of the two node embeddings was used to construct the edge embeddings,and the edge embeddings were put into the binary classifier based on the gradient boosting decision tree algorithm.Therefore,the link prediction problem in heterogeneous networks was solved.Experimental results show that the improved algorithm has higher prediction accuracy and stability.