【作者】 杜紫维；

【导师】 赵姝； 陈洁；

【作者基本信息】 安徽大学 ， 计算机科学与技术， 2021， 硕士

【摘要】 网络表示学习,又称为网络嵌入,旨在将网络数据嵌入到潜在的低维向量空间,同时保留网络的固有特性,如结构相似性,属性相似性等。节点在嵌入空间中的低维向量表示可以作为机器学习算法的输入,应用于大量的下游任务中,如节点分类,链路预测,具有丰富的实际应用价值。近年来网络表示学习引起研究者们的极大关注,在单粒度网络表示学习上取得了很大的成就。然而,现实世界中很多网络是呈现多粒度特性的,这一特性已被证明有助于完成一系列的网络分析任务。因此,如何快速地获得网络的多粒度结构并在学习的过程中保留这一特性成为网络表示学习领域的重要挑战。目前,网络表示学习方法主要从网络的单粒度局部结构(如一阶近似度、二阶近似度和社区结构)或者单粒度的属性信息来学习节点的表示,也有少数的学者对多粒度结构进行研究。本文结合粒计算思想,对多粒度网络表示学习方法进行研究,通过粒化过程构建多粒度网络可以简化网络规模,再通过已有的方法求原始网络的近似解,最后逐层细化得到原始网络的节点表示。但是现实世界的网络不仅具有网络结构,同时还具有丰富的节点属性信息且很多呈现多粒度结构。本文进一步的融合结构和节点的属性信息研究多粒度属性网络表示学习方法。本文的主要贡献如下:(1)提出一种基于多粒度结构的网络表示学习方法框架(HSNE)。该方法不仅能够保留网络的局部结构,还能在快速学习的过程中保留网络的多粒度结构。首先,根据网络结构之间的关系构造不同的粒,根据边构建粒层,从而将网络粒化为一个多粒度网络。然后在最粗层利用已有的网络表示学习方法快速学习原网络的近似解,根据不同粒层之间的关系,从粗到细逐层将近似解细化,最后获得原始网络的最终解,即每个节点的表示。实验结果表明提出的方法能够很好的保留多粒度结构特性。(2)提出一种快速的多粒度属性网络表示学习方法(HANE)来保留多粒度属性信息。具体来说,HANE设计一种融合结构和节点属性的粒化方法,且可以使用任何(仅基于结构的或结合属性信息的)无监督网络嵌入学习最粗网络的节点表示,最后融合节点的属性细化节点表示。HANE在保持网络表示学习性能的同时提高网络表示学习的速度,且最粗层网络表示学习方法非常灵活。本文在六个数据集和两个下游任务上对提出的框架HANE进行广泛的评估。实验结果表明,HANE算法在效率和有效性上比已有的网络嵌入算法有显著的改进。更多还原

【Abstract】 Network representation learning,also called network embedding,aiming to learn low dimensional vectors for nodes while preserving essential properties of the network,such as structural similarity,attribute similarity,etc.The low-dimensional vector of the node can be used as the input of the machine learning algorithm and applied to a lot of downstream tasks,such as node classification and link prediction,benefits plenty of practical applications.In recent years,network representation learning has attracted great attention from researchers,making great achievements on single-granularity network embedding.In the real world,many networks present multi-granularity structure,which has been shown to contribute to the completion of network analysis tasks.Therefore,how to quickly obtain the multi-granularity structure of the network and preserve the multi-granularity structure of the network is a meaningful and tough task in the field of network representation learning.Most existing methods are based on single-granularity,which learn representations from local structure of nodes(such as first-order proximity,second-order proximity,and community structure)or single-granularity attribute information.Multi-granularity network representation learning has been least studied.In this dissertation,we combine granular computing to research multi-granularity network representation learning methods.Constructing a multi-granularity network by granulation model,which can simplify the network scale.And then learn on the coarsest network to get the approximate solution of the original network.Finally,the approximate solution is refined from coarse to fine,and the final solution of the original network is obtained.However,the real-world network not only has a network structure,but also has rich node attribute information and many networks present multi-granularity structure.This work further studies the representation learning method of multi-granularity attribute network by fusing structure and node attribute information.The main work of this dissertation is as follows:(1)We propose a hierarchical structure network embedding framework(HSNE),which not only preserves the local structure,but also preserves the multi-granularity structure.First,construct different granules according to the relationship between the network structure,construct the granular layer according to the edges,thereby construct a multi-granularity network with a gradually decreasing network scale,and then learn on the coarsest network to get the approximate solution of the original network.Finally,according to the relationship between different granular layers,the approximate solution is refined from coarse to fine,and the final solution of the original network is obtained.Experimental results show that the proposed method can preserve the multi-granularity structure well.(2)We propose a hierarchical attributed network embedding framework HANE to preserve the multi-granularity attributed information.Specifically,for an attributed network,HANE designs a granulation method that combines structure and node attributes.After using any unsupervised network embedding method(such as structure-only network embedding or attributed network embedding)to learn node representations of the coarsest network.HANE refines the nodes representations of the hierarchical attributed network from coarse to fine.HANE improves the speed of network representation learning while maintaining its performance and the representation learning method of the coarsest network is flexible.We conduct extensive evaluations for the proposed framework HANE on six datasets and two benchmark applications.Experimental results demonstrate that HANE achieves significant improvements compared to the state-of-the-art network embedding methods in efficiency and effectiveness.更多还原