Scholarlink

<p>Graph contrastive learning (GCL) methods extract intrinsic semantic information from graph structures for recommendation purposes, achieving promising results. However, existing data augmentation strategies in graph contrastive learning may alter the intrinsic semantic information of the original interaction data, especially when it is susceptible to bias when disturbed by noise. Additionally, the over-smoothing problem in representation learning using GNN persists. To better extract and preserve the intrinsic semantic information of user–item interactions and alleviate the over-smoothing problem in graph neural network learning, a novel graph contrastive model (FDGCL) is proposed. By FDGCL, global information is extracted from original graph structure using principal component analysis and then fused with feature means of original graph structure to retain important interaction information for data augmentation. Both the original and augmented graph branches adopt GNNs for forward propagation and representation learning. Furthermore, an adaptive feature decorrelation loss is introduced to regulate GNN training by quantifying and minimizing the correlation between user and item features through correlation coefficients. Dynamically adjusting feature correlations enhances the model’s ability to capture personalized user preferences, improves representation capacity, and effectively mitigates the over-smoothing problem in GNNs. Experiments conducted on multiple datasets indicate that FDGCL outperforms current mainstream recommendation algorithms in terms of performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Feature decorrelation graph contrast learning based on PCA for recommendation

  • 期刊论文

  • 发表日期：

    2026年2月27日

查看全文
  • Zhi Liu,
  • Xincheng Xia,
  • Yunjie Huang,
  • Yihao Zhang,
  • Xiaojuan Liu

摘要

Graph contrastive learning (GCL) methods extract intrinsic semantic information from graph structures for recommendation purposes, achieving promising results. However, existing data augmentation strategies in graph contrastive learning may alter the intrinsic semantic information of the original interaction data, especially when it is susceptible to bias when disturbed by noise. Additionally, the over-smoothing problem in representation learning using GNN persists. To better extract and preserve the intrinsic semantic information of user–item interactions and alleviate the over-smoothing problem in graph neural network learning, a novel graph contrastive model (FDGCL) is proposed. By FDGCL, global information is extracted from original graph structure using principal component analysis and then fused with feature means of original graph structure to retain important interaction information for data augmentation. Both the original and augmented graph branches adopt GNNs for forward propagation and representation learning. Furthermore, an adaptive feature decorrelation loss is introduced to regulate GNN training by quantifying and minimizing the correlation between user and item features through correlation coefficients. Dynamically adjusting feature correlations enhances the model’s ability to capture personalized user preferences, improves representation capacity, and effectively mitigates the over-smoothing problem in GNNs. Experiments conducted on multiple datasets indicate that FDGCL outperforms current mainstream recommendation algorithms in terms of performance.