NodeHGAE: Node-oriented heterogeneous graph autoencoder

Heterogeneous graph autoencoder (HGAE), as an unsupervised learning approach, aims to encode nodes and edges of heterogeneous graphs into low-dimensional vector representations, and simultaneously reconstruct the original graph structure from node representations. Existing heterogeneous graph encode...

Ausführliche Beschreibung

Gespeichert in:

Bibliographische Detailangaben
Veröffentlicht in:	Information sciences Jg. 719; S. 122448
Hauptverfasser:	Zhu, Xiangkai, Li, Chao, Yan, Yeyu, Zhao, Zhongying, Duan, Hua, Zeng, Qingtian
Format:	Journal Article
Sprache:	Englisch
Veröffentlicht:	Elsevier Inc 01.11.2025
Schlagworte:	Graph neural network Heterogeneous graph autoencoder Heterogeneous graph representation learning Heterogeneous graph autoencoder Heterogeneous graph representation learning Graph neural network
ISSN:	0020-0255
Online-Zugang:	Volltext
Tags:	Tag hinzufügen Keine Tags, Fügen Sie den ersten Tag hinzu!

Beschreibung
Zusammenfassung:	Heterogeneous graph autoencoder (HGAE), as an unsupervised learning approach, aims to encode nodes and edges of heterogeneous graphs into low-dimensional vector representations, and simultaneously reconstruct the original graph structure from node representations. Existing heterogeneous graph encoders typically follow the metapath paradigm, encoding different semantic information and then employing decoders to reconstruct nodes attributes and edges information. However, the interaction between different semantic structures is underestimated which may lead to loss of semantic information. Moreover, employing graph-level unified attention mechanism to weigh the importance of different semantic structures of nodes is a suboptimal choice. Motivated by these challenges, a novel method named Node-oriented Heterogeneous Graph Autoencoder (NodeHGAE) is proposed. It first aggregates different semantic information based on node neighborhoods and utilizes the Chebyshev function to derive high-order neighborhood information of nodes. Then, low-rank matrix and parameter decoupling are proposed to assign node-specific attention and semantic information is integrated from different levels. Additionally, node-level and graph-level contrastive loss are proposed to redress the noise problem in the process of feature and topology coupling in HGAE. Experiments have shown that NodeHGAE outperforms state-of-the-art methods on four public heterogeneous graph datasets. The code of NodeHGAE can be found at Github.1
ISSN:	0020-0255
DOI:	10.1016/j.ins.2025.122448