基于局部建模的异构图表征学习方法

Heterogeneous graph representation learning via multi-relation ego-networks modeling

现有的异构图神经网络通常忽略了存在于邻居节点之间潜在的语义联系,这使得节点表征无法蕴含多样化的局部特征。为了解决这个问题,文中提出了一种基于局部建模进行表征学习的方法L2NH,该方法首先构建邻居节点之间的协方差矩阵,并通过关系编码将边的类型信息融入到矩阵中,然后基于局部网络特征值分解提出了一种多通道机制来使得节点表征蕴含多样化的局部特征,最后,实验表明,该方法能够有效地提升节点分类的效果。

The existing heterogeneous graph neural networks(HGNNs)ignore latent associations between neighbors and thus fail to reflect nodes’surrounding characteristics when performing neighbor aggregation.To tackle this issue,a novel embedding method based on multi-relation ego-networks modeling is proposed in this thesis.In particular,the covariance matrix of neighbor nodes is proposed to build dependencies between them while incorporating edge type information into the matrix by a simple but ef-fective relation encoding method.Then,a multi-channel mechanism based on the eigenvalue decomposition of ego-networks is pro-posed to enable node representations to reflect diverse global characteristics of ego-networks.Finally,extensive experiments demon-strate that the proposed method outperforms the state-of-the-art baselines in node classification task.