When graph neural networks meet deep nonnegative matrix factorization: An encoder and decoder-like method for community detection

Community detection is the best choice for accomplishing a variety of tasks in networks. It aims to find the optimal cluster structures, i.e., the high correlation between nodes in intra-cluster and low correlation between them in different clusters. Deep Nonnegative Matrix Factorization (DNMF) and...

Full description

Saved in:
Bibliographic Details
Published in:Expert systems with applications Vol. 271; p. 126676
Main Authors: Cheng, Junwei, He, Chaobo, Lin, Xuequan, Liu, Weixiong, Han, Kunlin, Tang, Yong
Format: Journal Article
Language:English
Published: Elsevier Ltd 01.05.2025
Subjects:
Autoencoder
Community detection
Deep nonnegative matrix factorization
Graph neural networks
Deep nonnegative matrix factorization
Community detection
Graph neural networks
Autoencoder
ISSN:0957-4174
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Community detection is the best choice for accomplishing a variety of tasks in networks. It aims to find the optimal cluster structures, i.e., the high correlation between nodes in intra-cluster and low correlation between them in different clusters. Deep Nonnegative Matrix Factorization (DNMF) and Graph Neural Networks (GNNs) have attracted much attention because of their advantages. And they have been widely adopted for community detection. Unfortunately, the performance of these methods is limited by their respective shortcomings. Therefore, their performance is distant from satisfactory. Especially, the innate feature of DNMF makes it difficult to cope with complex networks and withstand topology noise. Additionally, the performance of GNNs is limited by over-smoothing problem. To overcome the aforementioned flaws, we propose a novel method, SADNG, which connects the DNMF and GNNs seamlessly. Specifically, the over-smoothing problem of GNNs is alleviated when DNMF participates in convolution operations. And the abilities of DNMF to withstand topology noise and to reconstruct non-linear networks are boosted when GNNs participates in reconstructing weighted networks. In all, this way of cooperation compensates for their respective disadvantages. Experimental results across multiple benchmark datasets, including Cora and Citeseer, demonstrate that SADNG is superior to most state-of-the-art methods in community detection. •Integrates GCN and DNMF for enhanced community detection.•Addresses topology noise in dense networks effectively.•Mitigates over-smoothing in GCN using hierarchical mappings.•Outperforms state-of-the-art methods across multiple datasets.•Provides robust results in networks with hub nodes.
ISSN:0957-4174
DOI:10.1016/j.eswa.2025.126676