Scholarlink

<p>This paper presents an efficient solution to the hypergraph partitioning problem by introducing a novel multi-objective non-convex constrained model. We propose a new approach that significantly enhances partitioning quality and efficiency, employing the modified accelerated proximal gradient algorithm combined with the directional cosine-based weighted partitioning algorithm. To further improve partitioning results, we incorporate a parallel computation strategy that optimizes across multiple parameters and partitions, reducing the risk of falling into local optima. In the numerical experiments, the algorithm is compared with state-of-the-art partitioners such as KaHyPar, Mt-KaHyPar, and hMETIS on the ISPD98 and Titan23 benchmarks. The results show that although the proposed algorithm does not outperform the comparison partitioners in terms of running time, it exhibits a significant competitive advantage in partitioning quality. Specifically, in the weighted vertex partitioning task, the proposed algorithm successfully solves more than half of the instances, yielding the highest-quality solution among evaluated partitioners. In the Titan23 benchmarks (with unit weights), the algorithm solves 22.7% of the problems with the best-found solution, slightly outperforming KaHyPar.</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

Multi-objective k-way parallel hypergraph partitioning with proximal gradient algorithm

  • 期刊论文

  • 发表日期：

    2025年11月20日

查看全文
  • Yingying Li,
  • Hongwei Liu,
  • Hailong You,
  • Zexian Liu,
  • Fang Zhang

摘要

This paper presents an efficient solution to the hypergraph partitioning problem by introducing a novel multi-objective non-convex constrained model. We propose a new approach that significantly enhances partitioning quality and efficiency, employing the modified accelerated proximal gradient algorithm combined with the directional cosine-based weighted partitioning algorithm. To further improve partitioning results, we incorporate a parallel computation strategy that optimizes across multiple parameters and partitions, reducing the risk of falling into local optima. In the numerical experiments, the algorithm is compared with state-of-the-art partitioners such as KaHyPar, Mt-KaHyPar, and hMETIS on the ISPD98 and Titan23 benchmarks. The results show that although the proposed algorithm does not outperform the comparison partitioners in terms of running time, it exhibits a significant competitive advantage in partitioning quality. Specifically, in the weighted vertex partitioning task, the proposed algorithm successfully solves more than half of the instances, yielding the highest-quality solution among evaluated partitioners. In the Titan23 benchmarks (with unit weights), the algorithm solves 22.7% of the problems with the best-found solution, slightly outperforming KaHyPar.