.png){kind=icon}
.png){kind=icon}
Return
Diff-GNDCRec: A diffusion model with graph-node enhancement and difference comparison for recommendation
.png){kind=icon}
PRE
AI
DOI：10.1016/j.ipm.2025.104154.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Attention mechanism is widely used by Graph Neural Networks (GNNs) based recommender systems. However, data sparsity and noise can potentially disrupt the model, and consistence information within the graph structure may not fully capture the relative importance of graph nodes, which could influence the result of classification. Therefore, a diffusion model with graph-node enhancement and difference comparison for recommendation (Diff-GNDCRec) is proposed. Firstly, the feature vectors of the graph nodes are processed by Graph Convolutional Network (GCN) to obtain feature embedding, which is subsequently augmented by introducing Gaussian noise via a diffusion model. Secondly, augmented views of the graph nodes are generated through a multi-head Graph Attention Network (GAT) and denoised using average pooling to recover user interactions effectively. Finally, to better reflect the importance of the nodes, the model assigns weights to the nodes based on the neighborhood characteristics and combines the consistence and difference metrics to form the forward-supervised signals and contrast-supervised signals, respectively, and integrates them to improve the contrast learning effect. The model autonomously learns complex relationships between nodes, improving both recommendation accuracy and the model's generalization capability in a fuzzy environment. Comparative evaluations with eleven benchmark models across three real-world datasets- Tmall, Amazon, and Gowalla-demonstrate that Diff-GNDCRec improves recall and normalized discounted cumulative gain by 1.26% to 3.32% and 1.37% to 4.12%, respectively. These results demonstrate the effect of Diff-GNDCRec.
Keywords:
Diffusion model
Multigraph attention network
Similarity and difference contrast learning
Recommendation system
Journal
I
IF:
6.9
Papers:
5.2K
Citations:
1.4W