基于多粒度结构的网络表示学习

图卷积网络（GCN）能够适应不同结构的图，但多数基于GCN的方法难以有效地捕获网络的高阶相似性。简单添加卷积层将导致输出特征过度平滑并使它们难以区分，而且深层神经网络更难训练。本文选择将网络的多粒度结构和图卷积网络结合起来用于学习网络的节点特征表示，提出基于多粒度结构的网络表示学习方法Multi-GS。首先，基于模块度聚类和粒计算思想，用分层递阶的多粒度空间替代原始的单层网络拓扑空间；然后，利用GCN模型学习不同粗细粒度空间中粒的表示；最后，由粗到细将不同粒的表示组合为原始空间中节点的表示。实验结果表明:Multi-GS能够捕获多种结构信息，包括一阶和二阶相似性、社团内相似性（高阶结构）和社团间相似性（全局结构）。在绝大多数情况下，使用多粒度的结构可改善节点分类任务的分类效果。

Network representation learning based on multi-granularity structure

The Graph Convolution Network (GCN) can adapt to graphs with different structures. However, most GCN-based models have difficulty effectively capturing the high-order similarity of the network. Simply adding a convolution layer will cause the output features to be too smooth and difficult to distinguish. Moreover, the deep neural network is more difficult to train. In this paper, multi-granularity structure and a GCN are combined to represent the node characteristics of the learning network. A multi-granularity structure-based network representation learning method, Multi-GS, is proposed. First, based on the idea of modularity clustering and granular computing, hierarchical multi-granularity space was used to replace the original single-layer network topology space. The GCN model was then used to learn the representation of granules in different coarse- and fine-granularity spaces. Finally, representations of the different grains were combined into representations of nodes in the original space from coarse to fine. Experimental results showed that multi-GS can capture a variety of structural information, including first-order and second-order similarity, intra-community similarity (high-order structure), and inter-community similarity (global structure). In most cases, using multi-granularity structure can improve the classification performance of node classification tasks.