一种用于图卷积网络的社交关系方向门控算法

针对社交网络用户态度分析任务中用户之间原有社交关系方向可能阻碍态度信息流动以及标签扩散的问题，提出了一种应用于半监督图卷积网络的社交关系方向门控算法.该算法首先在原有与逆向社交关系方向上分别进行图卷积运算，得到2种用户节点态度特征向量，然后利用门控机制对2种特征向量进行动态融合.扩展了态度信息传播路径的同时，还能够捕捉用户影响力差异，以自动选择态度信息的流动方向.在2个真实热点话题数据集上的实验结果表明，现有图卷积网络在加入该算法之后，其用户态度分析的准确率能够得到有效提升.

A Social Relationship Direction Gating Algorithm for Graph Convolutional Networks

Facing the problem in social user attitude analysis that the original social relationship direction between users in social networks hinders the flow of attitude information and label propagation, a social relationship direction gating algorithm for graph convolutional networks is proposed. The algorithm first performs graph convolution on the origin and reverse social relationship directions to obtain two types of user node attitude feature vectors, and then leverages the gating mechanism to integrate the feature vectors dynamically. While expanding the propagation of attitude information, the algorithm can also capture the differences of user influence to automatically select the flow of attitude information. Experiments on two real hot topic datasets show that the accuracy of the existing graph convolutional networks can be effectively improved after adding the proposed algorithm.