This paper introduces a generalization of Convolutional Neural Networks (CNNs) from regular grid-structured data such as images to graphs with irregular linkage structures, especially expressive heterogeneous graphs with typed nodes and schemas. We propose a novel spatial convolution operation to model the key aspects of a CNN: local connectivity and translation invariance, using higher-order structures. We use the concept of motif, which provides us flexibility to describe spatial locality in Heterogeneous graphs with respect to its schema, or any graph with respect to desired higher-order connection patterns. We develop a novel neural network architecture Motif-CNN that captures higher-order structural and feature information by combining the information extracted from multiple patterns through deeper layers. Our experiments on semi-supervised learning tasks on social networks and real-world heterogeneous graphs indicate significant gains over existing graph CNNs and respective state-of-the-art techniques.

该论文介绍了一种新的卷积神经网络架构，名为Motif-CNN，能够处理具有不规则连接结构的图形，特别是异构图形，实现了半监督节点分类，并在真实社交网络和多个代表性异构图形数据集上显示出与现有图形卷积神经网络和其他最先进技术相比的显着优势。

基于模式的图卷积神经网络