.png){kind=icon}
.png){kind=icon}
Return
Diffusion-enhanced graph contrastive learning with hierarchical negative sampling for link prediction
.png){kind=icon}
PRE
AI
DOI：10.1016/j.knosys.2026.116025.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
As a fundamental task in network analysis, link prediction is important for practical applications such as intelligent recommendation systems and urban traffic flow forecasting. This paper proposes a Diffusion-Enhanced Graph Contrastive Learning with Hierarchical Negative Sampling (DGCN) for link prediction. Specifically, our DGCN introduces a diffusion model incorporating semantic transformation mechanisms to generate dual-contrast views with distinct semantic characteristics. This process is controlled via a Feature-wise Linear Modulation (FiLM) layer, which dynamically fuses time-step embeddings and query-node features to preserve semantic relevance during diffusion. It is capable of capturing local structural characteristics within a one-hop neighborhood while simultaneously extracting global semantic information across k -hop neighborhoods, where k is adaptively determined. This dual-view design effectively enables the collaborative representation of local fine-grained features and global semantic patterns. More importantly, this model introduces a dynamic time-step adjustment mechanism during the diffusion denoising process, which enables the automatic generation of negative samples with diverse semantic levels and finely calibrated difficulty classifications in the latent space. Based on these negative samples, we develop a contrastive learning framework aimed at improving model performance through the optimization of embedded representations from both the main view and the contrasting views. Experimental results on multiple public datasets demonstrate that our method consistently outperforms state-of-the-art approaches. Our datasets and source code are available.1
Keywords:
link prediction
graph contrastive learning
diffusion model
hierarchical negative sampling
semantic transformation
Journal
K
IF:
7.6
Papers:
1.2W
Citations:
4.5W