基于折叠路径聚合的属性网络节点嵌入方法

Node Embedding Method Based on Folded Path Aggregation on Attributed Network

属性网络嵌入是图分析领域具有挑战性的任务之一,旨在从网络的拓扑结构和节点特征中学习节点的低维向量表示,同时最大限度地保持其结构和固有特性,然而现有方法多数仅研究网络间的基本关系,未考虑节点邻居的相对重要性。基于此,提出一种基于折叠路径聚合的属性网络节点嵌入方法,有效挖掘属性网络中的复合关系并充分度量节点邻居的重要程度。基于拓扑结构捕获节点的直接邻居并构建结构-属性二部图,挖掘“节点-属性-节点”折叠路径中所蕴含的复合关系,捕获节点的间接邻居。设计语义路径内部聚合策略,通过卷积神经网络聚合器聚合间接邻居表示和直接邻居表示,同时融合节点属性以捕获不同语义之间细粒度的特征交互,并根据语义路径间聚合策略整合2种细粒度嵌入表示,得到最终的节点嵌入。在Flickr、ArXiv和Pubmed这3个真实数据集上的实验结果表明,该方法的节点分类性能优于先进的属性网络嵌入方法,且与Node2Vec方法相比,Macro-F1和Micro-F1值分别高出0.067~0.234。

Attributed network embedding is a challenging task in the field of graph analysis.It aims to learn the low-dimensional vector representation of nodes from the network topology and node attributes of the network while maintaining its structure and inherent characteristics to the greatest extent possible.However,most of the existing methods often only consider the basic relationship between networks and neglect the relative importance of node neighbors.To address this shortcoming,this paper proposes an attributed network node embedding method based on folding path aggregation,which aims to effectively mine the composite relationships in the attributed network and fully measure the importance of node neighbors.First,the direct neighbors of nodes are captured based on the topology,and the structure-attribute bipartite graph is constructed.Furthermore,the compound relationship present in the node-attribute-node folding path is mined to capture the indirect neighbors of the node.Second,an internal aggregation strategy for semantic paths is designed.The indirect-and direct-neighbor representations are aggregated using a convolution neural network aggregator,and the node attributes are fused to capture the fine-grained feature interaction between different semantics.The final embedding can be achieved by integrating two fine-grained embedding representations through the aggregation between semantic paths.Experimental results on three real-world datasets,Flickr,ArXiv and Pubmed,show that the node classification performance of the proposed method is better than that of the advanced attribute network embedding method,and the values of Macro-F1 and Micro-F1 are 0.067-0.234 higher than Node2Vec,respectively.