融合抗噪机制的图变分自编码器社交网络推荐方法研究

Research on Social Network Recommendation Method Based on Graph Variational Autoencoder with Noise-Resistant Mechanism

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,是社交网络推荐中使用的主要方法之一。随着社交网络近些年来蓬勃发展,数据量的剧烈增加无可避免地导致坏数据的出现(即目标节点特征属性缺失,或是目标节点特征属性错乱)。为了解决有节点属性社交网络中因为节点特征缺失导致的链路预测准确率降低的问题,提出了一种结合噪声对抗机制的图变分自编码器模型(Denoising Graph Variational Autoencoder,DGVAE)来优化链路预测效果。通过建立一种图变分自编码器结构,并设计一个噪声对抗模块,使得图变分自编码器能够有效地抵抗噪声干扰。经过一系列实验的验证,在有节点属性的网络中,采用噪声对抗的图变分自编码器模型能够有效地预测复杂的网络结构,而且在数据有冗余噪声的情况下,这种模型的预测效果有显著的改善。

Link prediction is to predict the possibility of connection between two non-adjacent nodes in a complex network based on the existing topological information in the network,and is one of the main methods used in social network recommendation.With the vigorous development of social networks in recent years,the dramatic increase in the amount of data inevitably leads to the emergence of bad data(that is,the lack of characteristic attributes of target nodes,or the disorder of characteristic attributes of target nodes).To address the problem of lower link prediction accuracy caused by the lack of node characteristics in social networks with node attributes,a DGVAE(Denoising Graph Variational Autoencoder)model is proposed to optimize the link prediction effect.By establishing a graph variational autoencoder structure and designing a noise resistance module,the graph variational autoencoder can effectively resist noise interference.After a series of experimental verification,in the network with node attributes,the graph variational autoencoder model using noise confrontation can effectively predict complex network structures,and in the case of data with redundant noise,the prediction of this model is significantly improved.