【作者】 李培森；

【导师】 王国胤；

【作者基本信息】 重庆邮电大学 ， 计算机科学与技术， 2021， 硕士

【摘要】 随着现代人工智能与智能信息服务的发展,大规模数据不断更新,各领域形成了复杂的信息网络。挖掘网络中的有效信息、准确的分析网络实体的性质,有利于各领域掌握行内市场动态、社会趋势等。网络表示学习作为复杂信息网络分析的基础,旨在将网络中海量稀疏数据表示为低维稠密的实值向量,以适用于链路预测、节点分类等实际任务。近年来,在分布式技术的支持下,大规模的复杂网络表示学习得到了发展。而现有的研究主要集中在单一的粗粒度上,即依赖于静态的网络拓扑关系,很少考虑网络的附加信息等外部知识,而丰富的外部知识能够为网络表示学习研究及其后续任务提供理论推导能力和可解释性。获取细粒度的外部知识和粗层面语义信息的关联性,并建立实体与网络节点的对应关系,直接决定了网络拓扑结构及性质,大幅度提升网络表示学习的关系推理与预测能力。因此,融合多种细粒度的附加信息以兼容的形式进行网络表示学习具有重要的意义。针对上述问题,本文结合多粒度认知的思想对网络表示学习进行深入研究,主要内容包括:1.提出了一种基于多粒度特征融合的网络表示学习算法,首先结合多粒度认知计算的思想,将复杂网络粒度细化,不仅考虑节点的拓扑结构,还包括丰富的节点内容,提出一种解决拓扑结构和属性信息关联性的融合机制,再通过深度学习技术得到统一的表示,保留节点潜在的相似度及语义信息。在下游的节点分类及链接预测验证实验中,提出的算法在真实的网络数据集上的性能和基准算法对比有较好的提升,能有效的进行网络表示学习。2.提出了一种基于注意力机制的多视图属性网络表示学习算法。该算法针对网络表示方法中属性信息的重要性程度不同以及全局的潜在信息保留不足的问题,提出一种多视图属性增强机制。首先,通过自注意力策略对一阶拓扑结构中邻居节点和具有相似社区附加属性按照注意力系数进行加权求和;同时,对每个节点的非邻居节点进行属性相似性度量,抽取一定数量的非直接关联的高阶节点继续进行有权重的属性加权,最后将两个视图进行融合,进一步提升表示效果。实验结果表面,算法有效的融合了网络结构和附加属性,提升了后续任务性能。更多还原

【Abstract】 With the development of modern artificial intelligence,intelligent information services and large-scale data are constantly updated,complex information networks have been formed in various fields.Mining the effective information in the network and analyzing the nature of the network entity are conducive to various fields to grasp the market dynamics and social trends in the industry.Network representation learning which has been regarded as the basis of complex information network analysis,aims to represent the massive sparse data in the network as low-dimensional dense real-valued vectors,which is suitable for practical tasks such as link prediction and node classification.With the support of distributed technology,large-scale complex network representation learning has been developed in recent years.Existing researches mainly focus on single coarse-grained,in another word,it relies on static network topology,which rarely consider external knowledge such as additional information of the network,rich external knowledge can provided to network representation learning research and subsequent tasks theory deduction ability and interpretability.Obtaining the correlation between fine-grained external knowledge and Coarse-grained semantic information,and establishing the corresponding relationship between entities and network nodes,directly determines the network topology and properties,which greatly improves the relational reasoning and prediction capabilities of network representation learning.Therefore,it is of great significance to integrate a variety of fine-grained external information and Coarse-grained semantic information to perform network representation learning in a compatible form.In response to the above issues,this article combines the idea of multi-granularity cognition to conduct in-depth research on network representation learning.The main contents include:1.A network representation learning algorithm based on multi-granularity information fusion is proposed.First,the algorithm combine the idea of multi-granularity cognitive computing to refine the granularity of complex networks.The algorithm not only considers the topological structure of the nodes,but also include the rich content of the nodes.A fusion mechanism of the relevance of structure and attribute information is proposed,which could obtain not only a unified representation through deep learning technology,but also the potential similarity and semantic information of nodes.In the downstream node classification and link prediction verification experiments,the performance of the proposed algorithm on the real network data sets better compared with other benchmark algorithms,and it can effectively perform network representation learning.2.A multi-view attribute network representation learning algorithm based on attention mechanism is proposed.In order to solve the issue that the different importance of attribute information in network representation methods and the insufficient retention of global potential information,this algorithm proposes a multi-view attribute enhancement mechanism.First,the self-attention strategy is used to perform weighted summation of the neighbor nodes and the additional attributes of similar communities in the first-order topology according to different weights.At the same time,the attribute similarity measurement is performed on the non-neighbor nodes of each node,and a certain number of high-level nodes are not directly related continue to perform weighted attribute weighting.Finally,the two views are fused to further improve the presentation effect.Experimental results show that the algorithm effectively integrates network structure and additional attributes to improve the performance of subsequent tasks.更多还原