基于多头注意力机制的社交网络符号预测

Social network sign prediction based on multi-head attention mechanism

传统符号预测方法缺少处理二阶邻居信息的能力,难以有效提取社交网络用户的低维特征。为了有效融合节点用户邻居信息,提出了一种利用多头注意力机制学习一阶、二阶邻居信息的网络表示学习方法(signed multi-head graph attention network,SMGAT),从而改善社交网络符号预测的效果。首先融合平衡理论和状态理论采样一阶邻居、二阶邻居;然后利用多头注意力机制融合邻居的符号和结构信息,学习节点的低维特征;最后通过逻辑回归分类器实现符号预测。通过在四个真实的符号网络数据集上进行实验,结果证明SMGAT方法能够有效挖掘邻居节点的符号和结构信息,提高社交网络符号预测效果。

Traditional sign prediction methods lack the ability to process the information of second-order neighbor,which are difficult to extract the users’low-dimensional features effectively.In order to effectively integrate the users’neighbor information,this paper proposed a signed network representation learning method(SMGAT)to improve the effect of social network sign prediction,which used multi-head attention mechanism to learn the first-order and second-order neighbor information.Firstly the method integrated the social balance theory and status theory and sampled the first-order neighbor and second-order neighbor.Then,it used the multi-head attention mechanism to integrate the neighbors’sign and structure information,learnt the low-dimensional features of nodes.Finally it realized the sign prediction through the logistic regression classifier.Through the experiments on four real signed networks,the results show SMGAT method can effectively mine the sign and structure information of neighbors,improve the performance of the social network link sign prediction.