融合节点结构和内容的网络表示学习方法

Network Representation Learning Method on Fusing Node Structure and Content

随着神经网络技术的快速发展,面向复杂网络数据的网络表示学习方法受到越来越多的关注,其旨在学习网络中节点的低维度潜在表示,并将学习到的特征表示有效应用于基于图的各种分析任务。典型的浅层随机游走网络表示学习模型主要基于节点结构相似和节点内容相似,不能同时有效捕获节点结构和内容的相似信息,因此在结构和内容等价混合的网络数据上表现较差。为此,探索了节点结构相似和节点内容相似的融合特征,提出了一种基于无监督浅层神经网络联合学习的表示方法SN2vec。实验分别利用节点结构和内容等价混合的Brazilian air-traffic,American air-traffic,Wikipedia数据集在多标签分类和降维可视化任务上进行验证。结果显示,SN2vec在多标签分类任务中的Micro-F1值优于现有的浅层随机游走网络表示方法,并且可以较好地学习到潜在结构表示一致的节点。

With the rapid development of neural network technology,the network representation learning method for complex network has got more and more attention.It aims to learn the low-dimensional potential representation of nodes in the network and to apply the learned characteristic representation effectively to various analysis tasks for graph data.The typical shallow random walk network representation model is mainly based on two kinds of characteristic representation methods,which are the node structure similarity and node content similarity.However,the methods can’t effectively capture similar information of node structure and content at the same time,and perform poorly on the network data with the equivalent structure and content.To this end,this paper explores the fusion characteristics of node structure and node content,and proposes a representation method called SN2vec,which is based on joint learning of unsupervised shallow neural networks.Further,in order to validate the effectiveness of the proposed model,this paper respectively conduct the multi-label classification and down-dimensional visualization tasks in Brazilian air-traffic,American air-traffic,and Wikipedia datasets.The results show that the Micro-F1 of using SN2vec in multi-label classification task is better than the existing shallow random walk network representation methods,and SN2vec can also learn better potential structural representation of consistent nodes.