Scholarlink

Abstract <p>Discovering causal relationships from observed data has applications in many fields including Bioinformatics. The PC (Peter and Clarke) algorithm and its parallel version are the state-of-the-art methods for causal discovery. However, the high runtime of these algorithms makes them inefficient for high-dimensional data. For example, to discover causal regulatory expression in gene, the number of nodes (factors) can easily be more than 2000. For such an instance, even the fastest variant of PC today would take days to complete. To solve this, we propose a recursive parallel causal discovery algorithm (RPCD). RPCD constructs the graph recursively, which leads to significant savings in the number of conditional tests. We evaluated RPCD on a number of causal network datasets as well as on a real-life incident dataset. The experimental results show that it improves both the running time and accuracy metrics significantly.</p> Graphical Abstract <p></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

A Fast Algorithm for High-Dimensional Causal Discovery

  • 期刊论文

  • 发表日期：

    2025年6月8日

查看全文
  • Sakib A. Mondal,
  • Prashanth Rv,
  • Sagar Rao

摘要

Abstract

Discovering causal relationships from observed data has applications in many fields including Bioinformatics. The PC (Peter and Clarke) algorithm and its parallel version are the state-of-the-art methods for causal discovery. However, the high runtime of these algorithms makes them inefficient for high-dimensional data. For example, to discover causal regulatory expression in gene, the number of nodes (factors) can easily be more than 2000. For such an instance, even the fastest variant of PC today would take days to complete. To solve this, we propose a recursive parallel causal discovery algorithm (RPCD). RPCD constructs the graph recursively, which leads to significant savings in the number of conditional tests. We evaluated RPCD on a number of causal network datasets as well as on a real-life incident dataset. The experimental results show that it improves both the running time and accuracy metrics significantly.

Graphical Abstract