Scholarlink

Graph Neural Networks (GNNs) have become increasingly essential in modeling complex clinical data, thereby facilitating heterogeneous data-based disease diagnosis. However, the application of GNNs to large-scale clinical data faces challenges due to their exponentially increasing computational costs and memory requirements, which restrict their effectiveness in medical image classification. Dividing large-scale graphs into subgraphs through partition methods emerges as a significant strategy for reducing computational resource consumption in graph learning. Nonetheless, this subgraph partition method requires traversing all subgraphs during training, significantly prolonging model convergence. To address these issues, in this study, we proposed a topology and embedding diversity-based sampling strategy, along with a hyperedge-based graph partition framework (DHSampling) to enhance the classification performance of subgraph-based GNNs. First, unlike traditional edges connecting only two nodes for each edge, we randomly assign nodes to hyperedges for forming a hypergraph, which connects multiple nodes simultaneously, allowing for the representation of complex relationships involving multiple entities. Then, we sample a subset of hyperedges with the highest diversity in both topology and embeddings to train the GNNs, providing accelerated training while maintaining minimal performance drops when sampling a subset for training. To the best of our knowledge, we are the first to utilize hyperedges in conjunction with diversity-based sampling to address the challenges faced by GNNs when applied to large-scale clinical data. Extensive experiments on two large-scale medical image classification benchmark datasets demonstrate that our DHSampling strategy can not only markedly reduce the model training time, but also achieve excellent classification performance compared to existing representative methods without increasing computational resource occupancy excessively. Our DHSampling code is available at https://github.com/basiralab/DHSampling .

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

DHSampling: Diversity-Based Hyperedge Sampling in GNN Learning with Application to Medical Imaging Classification

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Jiameng Liu,
  • Furkan Pala,
  • Islem Rekik,
  • Dinggang Shen

摘要

Graph Neural Networks (GNNs) have become increasingly essential in modeling complex clinical data, thereby facilitating heterogeneous data-based disease diagnosis. However, the application of GNNs to large-scale clinical data faces challenges due to their exponentially increasing computational costs and memory requirements, which restrict their effectiveness in medical image classification. Dividing large-scale graphs into subgraphs through partition methods emerges as a significant strategy for reducing computational resource consumption in graph learning. Nonetheless, this subgraph partition method requires traversing all subgraphs during training, significantly prolonging model convergence. To address these issues, in this study, we proposed a topology and embedding diversity-based sampling strategy, along with a hyperedge-based graph partition framework (DHSampling) to enhance the classification performance of subgraph-based GNNs. First, unlike traditional edges connecting only two nodes for each edge, we randomly assign nodes to hyperedges for forming a hypergraph, which connects multiple nodes simultaneously, allowing for the representation of complex relationships involving multiple entities. Then, we sample a subset of hyperedges with the highest diversity in both topology and embeddings to train the GNNs, providing accelerated training while maintaining minimal performance drops when sampling a subset for training. To the best of our knowledge, we are the first to utilize hyperedges in conjunction with diversity-based sampling to address the challenges faced by GNNs when applied to large-scale clinical data. Extensive experiments on two large-scale medical image classification benchmark datasets demonstrate that our DHSampling strategy can not only markedly reduce the model training time, but also achieve excellent classification performance compared to existing representative methods without increasing computational resource occupancy excessively. Our DHSampling code is available at https://github.com/basiralab/DHSampling .