Scholarlink

<p>Clustering is a crucial field, finding universal application across criminology, pathology, and urban planning. It often benefts from local structure, whether in the input domain or across samples. However, treating these signals separately can limit quality and scalability on high-dimensional, large datasets. We present ConstellationNet, a modular pipeline that couples CNN-derived spatial embeddings with GNN-based neighborhood aggregation on a residual kNN graph, and optionally performs graph-augmented inference by linking test queries to the training graph. ConstellationNet delivers state-of-the-art results in supervised classification on MNIST and CIFAR-10 and competitive performance in unsupervised clustering, with up to 10 times fewer parameters and markedly shorter training times. Due to its rapid training and powerful capabilities, ConstellationNet holds promise in fields such as epidemiology and medical imaging, allowing it to quickly train on new data and develop robust responses.</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

ConstellationNet: Reinventing Clustering Through GNNs

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Aidan Gao,
  • Junhong Lin

摘要

Clustering is a crucial field, finding universal application across criminology, pathology, and urban planning. It often benefts from local structure, whether in the input domain or across samples. However, treating these signals separately can limit quality and scalability on high-dimensional, large datasets. We present ConstellationNet, a modular pipeline that couples CNN-derived spatial embeddings with GNN-based neighborhood aggregation on a residual kNN graph, and optionally performs graph-augmented inference by linking test queries to the training graph. ConstellationNet delivers state-of-the-art results in supervised classification on MNIST and CIFAR-10 and competitive performance in unsupervised clustering, with up to 10 times fewer parameters and markedly shorter training times. Due to its rapid training and powerful capabilities, ConstellationNet holds promise in fields such as epidemiology and medical imaging, allowing it to quickly train on new data and develop robust responses.