Rethinking link prediction: A multi-scale graph masked autoencoder

In link prediction tasks, Graph Self-Supervised Learning (GSSL) has enormous potential for tackling the fundamental issue of sparse labels in real-world graph data. However, existing mainstream methods based on contrastive and generative approaches do not simultaneously consider hierarchical informa...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Neurocomputing (Amsterdam) Jg. 664; S. 132115
Hauptverfasser: Zhang, Guotai, Zuo, Enguang, Yan, Ziwei, Chen, Chen, Chen, Cheng, Lv, Xiaoyi
Format: Journal Article
Sprache:Englisch
Veröffentlicht: Elsevier B.V 01.02.2026
Schlagworte:
Graph masked autoencoder
Graph self-supervised learning
Link prediction
Multi-scale
Link prediction
Graph masked autoencoder
Multi-scale
Graph self-supervised learning
ISSN:0925-2312
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:In link prediction tasks, Graph Self-Supervised Learning (GSSL) has enormous potential for tackling the fundamental issue of sparse labels in real-world graph data. However, existing mainstream methods based on contrastive and generative approaches do not simultaneously consider hierarchical information from both local neighborhood information and global structural information. Therefore, we revisit these two approaches from a novel perspective and propose a multi-scale graph masked autoencoder (MS-GMAE) to enable the model to learn stable representations that capture aggregated local neighborhood and global structural information, thereby achieving precise predictions of both local and global links in link prediction tasks. Specifically, we propose a Cross-Correlation Decoder and a latent global representation prediction strategy. The Cross-Correlation Decoder utilizes multi-layer embedding representations of two nodes to decode and predict links between nodes, enabling the model to accurately perceive local neighborhoods. Latent Global Representation Prediction forces the reconstruction of embeddings that capture global structural information in the latent space, enhancing the model’s ability to capture global structural information. Additionally, we reconstruct stable contextual representations for nodes, enabling the model to learn more stable graph representation information. Results show that compared to 17 state-of-the-art baselines across five datasets, our model achieves state-of-the-art performance on four datasets, with the highest performance improvement reaching 9.83 %. •A multi-scale autoencoder learns hierarchical features for link prediction.•Cross-correlation decoding captures local patterns, mitigating overfitting.•Predicting latent global representations captures long-range dependencies.•Experimental results demonstrate superior performance over 17 baseline models.
ISSN:0925-2312
DOI:10.1016/j.neucom.2025.132115