基于改进图注意机制的网络嵌入方法研究及应用

网络已被广泛用作抽象现实世界系统以及组织实体之间关系的数据结构；网络嵌入模型是将网络中的节点映射为连续向量空间表示的强大工具；基于图卷积（Graph convolutional neural， GCN）的网络嵌入方法因受其模型迭代过程参数随机优化和聚合函数的影响，容易造成原始节点特征信息丢失的问题；为有效提升网络嵌入效果，针对于图神经网络模型在网络嵌入中节点表征学习的局限性，提出了一种基于二阶邻域基数保留策略的图注意力网络SNCR-GAT（Second-order Neighborhood Cardinality Retention strategy Graph attention network）,通过聚合二阶邻域特征基数的方式，解决网络节点潜在特征学习过程中重要信息保留问题；通过在节点分类和可视化两个网络嵌入应用任务上进行实验，结果表明，SNCR-GAT模型在网络嵌入上的性能表现相比较基准方法更具优越性。

Research and Application of Network Embedding Method Based on Improved Graph Attention Mechanism

Networks have been widely used as data structures for abstracting real-world systems and for organizing relationships between entities. The network embedding model is a powerful tool to map the nodes in the network into a continuous vector space representation. The network embedding method based on Graph convolutional neural (GCN) is easily affected by the random optimization of parameters in the model iteration process and the aggregation function. The problem of loss of original node feature information. In order to effectively improve the network embedding effect, a graph attention network based on the second-order neighborhood cardinality retention strategy is proposed for the limitation of the graph neural network model in the node representation learning in the network embedding. (SNCR-GAT, Second-order Neighborhood Cardinality Retention strategy Graph attention network), by aggregating the second-order neighborhood feature cardinality, it solves the problem of important information retention in the process of latent feature learning of network nodes; by classifying and visualizing two networks in nodes Experiments are carried out on the actual task of embedding, and the results show that the performance of the SNCR-GAT model on network embedding is more superior than the baseline method.