面向受损网络嵌入的深度降噪自编码器模型  

Deep Denoising Autoencoder Model for Embedding Damaged Networks

作　　者：李智杰[1] 王启辉 李昌华[1] 张颉[1] LI Zhi-jie;WANG Qi-hui;LI Chang-hua;ZHANG Jie(College of Information and Control Engineering,Xi′an University of Architecture and Technology,Xi′an 710055,China)

机构地区：[1]西安建筑科技大学信息与控制工程学院,西安710055

出　　处：《小型微型计算机系统》2022年第12期2535-2540,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61373112,51878536)资助;陕西省自然科学基金项目(2020JQ-687)资助。

摘　　要：针对现有结构深度网络嵌入模型在处理受损网络时,得到网络表示不能很好反映网络本质特征的局限,以及传统无监督降噪自编码器单层模型结构无法处理高度非线性复杂网络的问题,提出了结构深度降噪网络嵌入模型.通过引入掩盖噪声将受损网络邻接矩阵中的部分单元置零,以此作为模型输入,使用拉普拉斯特征映射处理相邻顶点,捕获网络的一阶相似度;将多个降噪自编码器堆叠得到深度降噪自编码器,利用深度降噪自编码器的多个非线性函数将输入数据映射到潜在空间并得到重构矩阵,最小化重构误差,以此来捕获网络的二阶相似度;联合优化一阶、二阶相似度来保持网络的局部和全局特征.在社交网络和语言网络上进行训练和测试,采用precision@k和MAP来评估模型性能.实验结果表明:相较于现有的网络嵌入模型,该模型实现了更优的网络重构与链路预测性能.In view of the limitations of the existing deep network embedding model when dealing with damaged networks,the network representation cannot reflect the inherent characteristics of the network,and the traditional unsupervised denoising autoencoder single-layer model structure cannot handle the problem of highly nonlinear and complex networks,The embedding model of structural depth noise reduction network is proposed.By introducing masking noise,some units in the adjacency matrix of the damaged network are set to zero,which is used as model input,and Laplacian feature mapping is used to process adjacent vertices to capture the first-order similarity of the network;multiple noise reductions are self-encoded The deep noise reduction autoencoder is stacked by the deep noise reduction autoencoder,and the multiple nonlinear functions of the deep noise reduction autoencoder are used to map the input data to the latent space and obtain the reconstruction matrix to minimize the reconstruction error to capture the second order of the network.Similarity:Joint optimization of the first-order and second-order similarity to maintain the local and global characteristics of the network.Train and test on social networks and language networks,and use precision@k and MAP to evaluate model performance.The experimental results show that compared with the existing network embedding model,this model achieves better network reconstruction and link prediction performance.

关 键 词：网络嵌入 深度学习 降噪自编码器 网络重构 链路预测 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]