Scholarlink

<p>Attributed representation learning has proven pivotal in deciphering the intricate relationships inherent in graph-structured data by leveraging both node attributes and graph topology. However, its effectiveness often hinges on data quality, which is frequently compromised by sparsity or missing information. Two key challenges persist: (1) attribute sparsity hinders models from effectively capturing global and local patterns, limiting generalization; and (2) existing methods struggle to jointly integrate attribute and structural information, reducing predictive accuracy. To address these limitations, we propose a Boosting Sparse Attribute Learning Model with Optimized Vector Quantization (<b>Blossom</b>). Blossom employs a High-order Attribute Enhancement (HAE) module that leverages cross-attention to capture rich high-order attribute interactions, improving their alignment with structural information. Furthermore, an Optimized Vector Quantization (OVQ) module encodes attribute embeddings into discrete codes, enhancing computational efficiency and representation robustness. Finally, we introduce an auxiliary variable for graph reconstruction, which enriches the latent space with high-order and global structural information. Extensive experiments on seven benchmark datasets demonstrate that Blossom consistently outperforms existing approaches in downstream classification tasks. These results highlight Blossom’s strong capacity to handle attribute sparsity and its effectiveness in robust attributed representation learning.</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

Blossom: Boosting sparse attribute learning with optimized vector quantization for graph representation learning

  • 期刊论文

  • 发表日期：

    2025年10月29日

查看全文
  • Huizhi Li,
  • Xian Mu,
  • Dagang Li

摘要

Attributed representation learning has proven pivotal in deciphering the intricate relationships inherent in graph-structured data by leveraging both node attributes and graph topology. However, its effectiveness often hinges on data quality, which is frequently compromised by sparsity or missing information. Two key challenges persist: (1) attribute sparsity hinders models from effectively capturing global and local patterns, limiting generalization; and (2) existing methods struggle to jointly integrate attribute and structural information, reducing predictive accuracy. To address these limitations, we propose a Boosting Sparse Attribute Learning Model with Optimized Vector Quantization (Blossom). Blossom employs a High-order Attribute Enhancement (HAE) module that leverages cross-attention to capture rich high-order attribute interactions, improving their alignment with structural information. Furthermore, an Optimized Vector Quantization (OVQ) module encodes attribute embeddings into discrete codes, enhancing computational efficiency and representation robustness. Finally, we introduce an auxiliary variable for graph reconstruction, which enriches the latent space with high-order and global structural information. Extensive experiments on seven benchmark datasets demonstrate that Blossom consistently outperforms existing approaches in downstream classification tasks. These results highlight Blossom’s strong capacity to handle attribute sparsity and its effectiveness in robust attributed representation learning.