Scholarlink

<p>Federated learning (FL) involves clients independently collecting and training data. Discrepancies in client computational capabilities and data volume can result in waiting time, hindering convergence. Intuitively, employing pruning to optimize sample quantity can enhance training efficiency. However, existing works often overlook this efficiency improvement due to the impact of the stage of pruning execution and the amount of pruned data on performance. To address this, we propose an Adaptive Data Pruning (AdaDP) method to dynamically adjust the sample size for heterogeneous clients. The key insight of AdaDP is that the initial FL phase is crucial for achieving optimal performance, as most samples can be learned quickly and then disregarded without affecting the model. Therefore, we formulate adaptive pruning as an optimization problem constrained by the initially obtained client response times to calculate the sample size. Experimental results demonstrate that AdaDP effectively balances efficiency and model performance, outperforming other methods.</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

Adaptive data pruning for efficient federated learning on heterogeneous clients

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Keke Yang,
  • Wei Fan,
  • Zheyuan Shen,
  • Xiangyu Fan,
  • Jing Li

摘要

Federated learning (FL) involves clients independently collecting and training data. Discrepancies in client computational capabilities and data volume can result in waiting time, hindering convergence. Intuitively, employing pruning to optimize sample quantity can enhance training efficiency. However, existing works often overlook this efficiency improvement due to the impact of the stage of pruning execution and the amount of pruned data on performance. To address this, we propose an Adaptive Data Pruning (AdaDP) method to dynamically adjust the sample size for heterogeneous clients. The key insight of AdaDP is that the initial FL phase is crucial for achieving optimal performance, as most samples can be learned quickly and then disregarded without affecting the model. Therefore, we formulate adaptive pruning as an optimization problem constrained by the initially obtained client response times to calculate the sample size. Experimental results demonstrate that AdaDP effectively balances efficiency and model performance, outperforming other methods.