Search

Community structure is an important characteristic of complex networks. It determines where important functions of a network are located. Recently, discovering community structure in complex networks has become a hot topic of research. However, the continuous increase in network size has made network structure more complex, and community detection has become extremely difficult in real applications. In particular, the detection results are usually not accurate enough when classical clustering methods are applied to high-dimensional data matrices. In this paper, inspired by the relationship between vertices, we design a novel and effective network adjacency matrix transformation method to describe vertices' similarity in the network topology. On this basis, we propose a framework to extract nonlinear features: community detection with deep transitive autoencoder (CDDTA). This framework can obtain powerful nonlinear features of a real network to make community detection algorithms perform excellently in practice. We further incorporate unsupervised transfer learning into the CDDTA (Transfer-CDDTA) by minimizing the Kullback–Leibler divergence of embedded instances, to discover powerful low-dimensional representations. Finally, we propose a new training strategy and optimization method for our algorithm. Extensive experimental results indicate that our new framework can ensure good performance on both real-world networks and artificial benchmark networks, which outperforms most of the state-of-the-art methods for community detection in social networks. [ABSTRACT FROM AUTHOR]

When analyzing graph-structured data, graph neural networks (GNNs) use well-designed propagation mechanism to extract node features. However, these methods only incorporate data from nodes' fixed neighborhoods or whole graph, not accounting for the influence of important nodes outside of the neighborhoods. We offer transferable graph neural networks with deep alignment attention (TGDAA), which is a novel neural network framework that designs aligned attention to assign different weights to related important nodes. In such a scenario, we concurrently remove a number of graph neural network obstacles and make TGDAA accessible for transfer learning issues. Extensive experiments in four aspects: traditional metrics, attention-related methods, varying label rates and ablation study show that the suggested TGDAA performs better than ten state-of-the-art techniques on ten benchmark network datasets. [ABSTRACT FROM AUTHOR]