Scholarlink

<p>Spectral clustering algorithms have been used in various research domains to discover structure and patterns in data. However, high computational and space complexity hinders their usage for large-scale datasets in machine learning and bioinformatics. Various approximate spectral clustering methods were proposed in the open literature to solve those problems. In this paper, we describe our GPU-based, parallel implementation of an approximate spectral algorithm based on the Nystrom method and column sampling and its memory-efficient variant. We evaluate our solution using several annotated datasets, such as USPS, MNIST, and MNIST8, as well as bioinformatics data, especially from the domain of single-cell and spatial transcriptomics. We obtain speedups of up to 31.8x depending on the dataset used and demonstrate the scalability of the solution for the datasets with up to four million samples.</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

Fast Parallel CPU-GPU Approximate Spectral Clustering for Transcriptomics Data

  • 期刊论文

  • 发表日期：

    2025年1月30日

查看全文
  • Stefan Branković,
  • Lazar Smiljković,
  • Predrag Obradović,
  • Miloš Radonjiić,
  • Marko Mišić

摘要

Spectral clustering algorithms have been used in various research domains to discover structure and patterns in data. However, high computational and space complexity hinders their usage for large-scale datasets in machine learning and bioinformatics. Various approximate spectral clustering methods were proposed in the open literature to solve those problems. In this paper, we describe our GPU-based, parallel implementation of an approximate spectral algorithm based on the Nystrom method and column sampling and its memory-efficient variant. We evaluate our solution using several annotated datasets, such as USPS, MNIST, and MNIST8, as well as bioinformatics data, especially from the domain of single-cell and spatial transcriptomics. We obtain speedups of up to 31.8x depending on the dataset used and demonstrate the scalability of the solution for the datasets with up to four million samples.