计算机科学与探索 ›› 2021, Vol. 15 ›› Issue (7): 1279-1288.DOI: 10.3778/j.issn.1673-9418.2006069

• 人工智能 • 上一篇    下一篇

层次标签引导的属性网络嵌入

陈洁,陈嘉琳,赵姝,张燕平   

  1. 1. 安徽大学 计算智能与信号处理教育部重点实验室,合肥 230601
    2. 安徽大学 计算机科学与技术学院,合肥 230601
  • 出版日期:2021-07-01 发布日期:2021-07-09

Hierarchical Labels Guided Attributed Network Embedding

CHEN Jie, CHEN Jialin, ZHAO Shu, ZHANG Yanping   

  1. 1. Key Laboratory of Intelligent Computing and Signal Processing, Ministry of Education, Anhui University, Hefei 230601, China
    2. School of Computer Science and Technology, Anhui University, Hefei 230601, China
  • Online:2021-07-01 Published:2021-07-09

摘要:

网络嵌入是在保持网络性质不变的前提下,将节点转换为低维向量,以便下游任务的求解。现有网络嵌入方法的研究大多关注于网络结构、节点属性信息或单层次标签信息等方面。然而,许多真实世界的网络节点通常具有丰富的层次标签信息,这些层次标签信息对获取高效的网络嵌入具有重要价值。由于不同层次的标签之间的信息很难相互关联或继承,如何合理利用层次标签信息进行网络嵌入,获得更高效的向量表示是亟待研究的问题。针对上述问题,提出了一种新的基于层次标签的属性网络嵌入框架(HLANE),该框架利用层次注意力机制将层次标签信息融入网络嵌入中。HLANE框架首先通过现有的网络嵌入方法获取结构和/或属性信息初始化节点的嵌入向量。然后通过层次注意力机制层建立多层次标签的父节点和子节点之间的联系,并依此指导网络节点初始化嵌入向量在不同层次的学习,最终生成网络节点的多层次嵌入向量表示。在真实数据集上的实验表明,与对比算法相比,HLANE框架具有更好的网络节点嵌入表示。

关键词: 网络嵌入, 属性网络, 层次标签, 层次注意力机制

Abstract:

Network embedding, aiming to learn low dimensional vectors for nodes while preserving important properties of the network, benefits plenty of network applications. Most existing works focus on network structure, node attribute information or label information. However, many real world networks are often associated with abundant hierarchical labels information, which is potentially valuable in seeking more effective network embedding. Since the information between labels in different levels is hard to correlate or inherit, how to make reasonable use of hierarchical label information to learn more efficient network embedding is still an urgent problem. To address the above issues, a novel hierarchical labels guided attributed network embedding framework (HLANE) is proposed. This framework incorporates hierarchical labels information into network embedding with the help of hierarchical attention layer. HLANE first captures structure and/or attributes information by any existing network embedding method to initialize embeddings. Then the hierarchical attention layer, which builds the connection between the parent labels and the child labels, integrates the hierarchical labels information to guide initial embedding so that it generates hierarchical embedding and entire embedding results with hierarchical labels information. Experiments on real-world datasets demonstrate that the proposed method achieves significantly better performance compared with the state-of-the-art embedding algorithms.

Key words: network embedding, attributed network, hierarchical labels, hierarchical attention