计算机科学
自编码
人工神经网络
节点(物理)
鉴定(生物学)
图形
人工智能
理论计算机科学
植物
结构工程
工程类
生物
作者
You Xiong,Zheng Hu,Chang Su,Shi‐Min Cai,Tao Zhou
标识
DOI:10.1016/j.asoc.2024.111895
摘要
Identifying vital nodes in complex networks is one of the critical research problems in network science. The emergence of deep learning has spurred the development of numerous approaches for vital node detection, particularly those leveraging graph embedding techniques. However, current methodologies often rely on partial network information to create latent representations of nodes, neglecting critical topological aspects of graphs. In this paper, we propose a novel fusion model, named AGNN, which merges Autoencoder and Graph Neural Network (GNN) architectures to fully integrate the structural details of networks into graph embeddings, thereby enhancing the model's efficacy. Initially, a graph convolutional network-based Autoencoder is introduced to produce latent representations of nodes that comprehensively incorporate the network's topological characteristics. Subsequently, AGNN uses the listMLE as the loss function to optimize the overall rank prediction model by combining the GNN and Susceptible-Infected-Recovered models. AGNN is trained on the synthetic network generated by the Barabási-Albert (BA) model without loss of generality and is then transferred to the real-world networks for vital node identification. We conduct extensive experiments on fifteen real-world network datasets in three metrics. AGNN has the highest average Kendall's τ coefficient at 0.7059, 5% above the second best. Additionally, it achieves the highest average monotonicity index of 0.9977, outperforming the runner-up at 0.9843. The experimental findings affirm the effectiveness of AGNN compared to benchmark and state-of-the-art techniques.
科研通智能强力驱动
Strongly Powered by AbleSci AI