基于局部图互信息最大化的异构图神经网络方法

Local graphical mutual information maximization based heterogeneous graph neural network

针对现有的基于互信息最大化的异构图神经网络(HGNN)方法因图读出操作的单射限制、粗粒度的特征保留而无法适用于现实网络的问题,提出一种基于局部图互信息最大化的、无监督的异构图神经网络方法。该方法使用元路径对异构图中涉及到的语义关系进行建模,并利用图卷积模块和语义级别的注意力机制来捕获单个节点的局部表征。该方法通过最大化单个节点与局部子图间的互信息,有效地学习高阶节点表征。实验结果表明,该方法相比基于全局图互信息的方法,可以将数据集DBLP/IMDB上的节点分类任务的微值F1(micro-F1)提高大约3%/9%,同时将DBLP/IMDB上的节点聚类任务的调整兰德系数(ARI)提高约23%/46%。

Aiming at the shortcomings of the injective ability of readout function and coarse-grained feature preservation in traditional mutual information maximization based heterogeneous graph neural networks( HGNN),which make them inadequate to use in the real-work networks,a new local graphical mutual information maximization based unsupervised heterogeneous graph neural network is presented.The model uses the meta-path to model the structure involving semantics in heterogeneous graphs and utilizes graph convolution module and semantic-level attention mechanism to capture individual node local representations.By maximizing the mutual information between the individual node embedding and the local graph,the proposed model effectively learns high-level node representations.The experimental results show that compared with HDGI which is based on the global graph mutual information maximization,the proposed method can increase the micro-F1 of the node classification task on DBLP/IMDB up to about 3%/9%,as well as the adjusted Rand index(ARI) of the node clustering task on DBLP/IMDB up to about 23%/46%.