【KDD 2019】Heterogeneous Graph Neural Network

背景

Representation learning in heterogeneous graphs aims to pursue a meaningful vector representation for each node so as to facilitate downstream applications such as link prediction, personalized recommendation, node classification, etc. 

异构图表示学习的目的是为每个节点寻找一个有意义的向量表示，以便于后续应用，如链路预测、个性化推荐、节点分类等。

This task, however, is challenging not only because of the demand to incorporate heterogeneous structural (graph) information consisting of multiple types of nodes and edges, but also due to the need for considering  heterogeneous attributes or contents (e.д., text or image) associated  with each node.

然而，这项任务具有挑战性，不仅是因为需要合并由多种类型的节点和边缘组成的异构结构（图）信息，而且还因为需要考虑与每个节点相关联的异构属性或内容（例如，文本或图像）。

Traditionally, a variety of these HetG tasks have relied on feature vectors derived from a manual feature engineering tasks. This
requires specifications and computation of different statistics or properties about the HetG as a feature vector for downstream machine learning or analytic tasks.

传统上，这些HetG任务依赖于从手工特征工程任务中派生的特征向量。这个 需要对HetG的不同统计或属性进行规范和计算，作为下游机器学习或分析任务的特征向量。

贡献

We formalize the problem of heterogeneous graph representation learning which involves both graph structure heterogeneity and node content heterogeneity.

我们将异构图表示学习问题形式化，该问题同时涉及图结构异构性和节点内容异构性。

We propose an innovative heterogeneous graph neural network model, i .e., HetGNN, for representation learning on HetG. Het-GNN is able to capture both structure and content heterogeneity and is useful for both transductive and inductive tasks. Table 1 summarizes the key advantages of HetGNN, comparing to a number of recent models which include homogeneous, heterogeneous,attributed graph models, and graph neural network.

我们提出了一种新颖的异构图神经网络模型，即HetGNN，用于HetG上的表示学习。Het-GNN能够同时捕获结构和内容的异构性，对于传导和归纳任务都是有用的。表1总结了HetGNN的主要优势，与一些最新的模型相比，这些模型包括同质、异构、属性图模型和图神经网络模型。

We conduct extensive experiments on several public datasets and our results demonstrate the superior performance of HetGNN over state-of-the-art baselines for numerous graph mining tasks including link prediction, recommendation, node classification & clustering, and inductive node classification & clustering.

我们在多个公共数据集上进行了大量的实验，结果表明HetGNN在许多图挖掘任务（包括链接预测、推荐、节点分类和聚类以及归纳节点分类和聚类）中的性能优于最先进的基线。

形式化定义

框架