一种降噪自编码器的复杂网络链路预测算法  

Link Prediction Algorithm Based on Denoising Autoencoder in Complex Networks

作　　者：曹志威[1] 樊志杰[1,2] 王青杨 韩伟力 李欣[4] CAO Zhi-wei;FAN Zhi-jie;WANG Qing-yang;HAN Wei-li;LI Xin(Department of Information Security Technology,The Third Research Institute of the Ministry of Public Security,Shanghai 200120,China;Software School,Fudan University,Shanghai 201203,China;College of Computer Science,Sichuan University,Chengdu 610207,China;School of Information Technology and Network Security,People′s Public Security University of China,Beijing 100038,China)

机构地区：[1]公安部第三研究所信息安全技术部,上海200120 [2]复旦大学计算机科学技术学院,上海201203 [3]四川大学计算机学院,成都610207 [4]中国人民公安大学信息网络安全学院,北京100038

出　　处：《小型微型计算机系统》2023年第3期665-672,共8页Journal of Chinese Computer Systems

基　　金：中国博士后科学基金项目(2020M670998)资助;上海市自然科学基金项目(21ZR1422000)资助;上海市人才发展资金项目(2020016)资助;四川省科技计划项目(2021YFS0310)资助。

摘　　要：链路预测是根据复杂网络中已有的拓扑信息预测网络中两个不相邻的节点间产生连接的可能性,是复杂网络领域中的重要研究方向,具有重要的研究价值.在理论层面上,提升链路预测算法的性能有利于更合理的挖掘和分析网络的演化机制;在应用层面上,提升链路预测算法的性能有助于补全网络拓扑的缺失信息,从而便于优化后续网络拓扑相关的算法,例如图表示学习和个性化推荐等.该领域尽管近些年已经取得了较多的研究成果,但依然存在不少缺陷.例如,作为主流的基于节点相似性的链路预测算法存在高度退化的问题,即对于大多数不相邻的节点对均输出相同的预测值;其次,由于不同的复杂网络在网络结构、节点度数、连边数量以及联通性上各有差异,然而当前的算法通常仅考虑网络的某种结构特征,因此只对于特定的网络类型预测效果较好,可扩展性较差.鉴于此,本文利用深度学习理论善于挖掘各种高维数据的重要特征,将无监督训练方法引入到复杂网络的链路预测中,提出一种基于降噪自编码器的复杂网络链路预测算法.该算法通过神经网络结构与损失函数的构造,首先使其具有数据降噪恢复的能力,然后将完整的训练集数据输入到模型中,即可实现预测复杂网络演化机制的目的.具体地,将加入噪声的邻接矩阵以列向量的方式逐条输入到神经网络结构中,然后运用该降噪自编码器模型确保输出向量与未加噪声的数据相近.经过反复训练,本模型中神经网络的结构和参数会不断调节,使其逐渐具备从低维数据中恢复高维信息的目的,进而达到预测复杂网络演化结构的效果.同时,该算法不仅能够从残缺数据中学习出有用的预测信息,而且能够降低复杂网络结构的差异性对算法的影响.通过在7种不同类型网络中的对比实验,分析结果表明本算法与其他经典的链路�Link prediction refers to the possibility of future connection between two nodes based on the existing topology information in the network. Link prediction is an important research direction in the field of complex networks, and it is of great value in theory and application. On the theoretical level, improving the prediction ability of link prediction algorithm is helpful to mining and analyzing the evolution mechanism of complex networks. On the application level, improving the prediction ability of link prediction algorithm is helpful to complete the missing topology information, so as to optimize subsequent topology-related algorithms, such as graph representation learning algorithm and personalized recommendation algorithm. In the field of link prediction, the researchers have achieved more achievements, but there are still many defects. For example, the link prediction algorithms based on node similarity, as a kind of mainstream algorithm, are often highly degraded, that is, most pairs of non-adjacent nodes have the same predicted values and are indistinguishable. Secondly, the different complex networks have the different structures, such as node degree, number of edges, and connectivity. However, many existing link prediction algorithms only consider certain structural characteristics of the network, so they are not well-suited for different types of networks, and the prediction effect is not stable. In term of this, and considering that deep learning theory is good at mining the important features of various high-dimensional data, an unsupervised training method is introduced into the link prediction of complex networks, and proposes a link prediction algorithm based on denoising autoencoder. Thus, the missing link information is recovered and predicted by mining the potential characteristics of network structure. Through the construction of neural network structure and loss function, the algorithm first has the ability of data denoising and recovery, and then inputs the complete training set data into t

关 键 词：复杂网络 复杂系统 链路预测 深度学习 无监督训练学习 降噪自编码器 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]