Scholarlink

<p>Graph representation learning on large-scale networks with high-dimensional node attributes poses significant computational challenges. Existing embedding methods often encounter a trade-off between scalability and the preservation of fine-grained structural and feature information, particularly when processing dense feature vectors. To address these limitations, this paper introduces GRACE (Graph-based Random Walk Autoencoder Compression for Embedding), a unified framework designed to optimize both representation quality and resource efficiency. The proposed architecture leverages a hierarchical autoencoder to perform dimensionality reduction on node features, minimizing computational overhead while retaining essential attribute semantics. This is coupled with an adaptive random walk mechanism that dynamically navigates the graph structure based on feature similarity to capture long-range dependencies. Furthermore, GRACE incorporates a graph attention mechanism to selectively aggregate neighborhood information, effectively mitigating the over-smoothing problem. Extensive experiments on benchmark datasets demonstrate that GRACE achieves superior node classification accuracy compared to state-of-the-art baselines, while significantly optimizing memory utilization and training time.</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

Grace: a scalable framework for graph embedding with autoencoder-based feature compression and random walks

  • 期刊论文

  • 发表日期：

    2026年4月7日

查看全文
  • Mohammad Mehdi Keikha,
  • Saman Barahoie,
  • Abolfazl Nadi

摘要

Graph representation learning on large-scale networks with high-dimensional node attributes poses significant computational challenges. Existing embedding methods often encounter a trade-off between scalability and the preservation of fine-grained structural and feature information, particularly when processing dense feature vectors. To address these limitations, this paper introduces GRACE (Graph-based Random Walk Autoencoder Compression for Embedding), a unified framework designed to optimize both representation quality and resource efficiency. The proposed architecture leverages a hierarchical autoencoder to perform dimensionality reduction on node features, minimizing computational overhead while retaining essential attribute semantics. This is coupled with an adaptive random walk mechanism that dynamically navigates the graph structure based on feature similarity to capture long-range dependencies. Furthermore, GRACE incorporates a graph attention mechanism to selectively aggregate neighborhood information, effectively mitigating the over-smoothing problem. Extensive experiments on benchmark datasets demonstrate that GRACE achieves superior node classification accuracy compared to state-of-the-art baselines, while significantly optimizing memory utilization and training time.