Scholarlink

This paper introduces autokRF, a family of automated learning algorithms designed to optimize hyper-parameters for k local Random Forest models in large-scale classification tasks. The autokRF algorithms automatically determine an appropriate number of clusters k to partition the training dataset, then train separate Random Forest models within each cluster. This localized approach enables efficient parallel classification on multi-core CPUs. To improve model performance, each cluster applies the .632 bootstrap estimator, while hyper-parameter tuning is carried out using grid search (GS), reinforcement learning (RL), and Bayesian optimization (BO). Experiments on the ImageNet dataset (1,281,167 images, 768 features, 1,000 classes) demonstrate that our proposal (kRF, autokRF-GS, autokRF-RL, and autokRF-BO) consistently outperform the standard Random Forest algorithm, achieving significant improvements in both classification accuracy and training efficiency. On a Linux Ubuntu 24.04 system with an Intel Core i5-12400 CPU (4.4 GHz, 6 cores, 12 threads) and 32 GB of RAM, these algorithms complete the ImageNet classification task in 9.2, 10.83, 11.04, and 10.58 min respectively, with classification accuracy over 88%.

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

AutokRF: Efficient and Scalable Local Ensemble Learning with Automated Hyper-Parameter Tuning for Large-Scale Classification

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Minh-Trung Nguyen,
  • Thanh Ma,
  • Thanh-Nghi Do

摘要

This paper introduces autokRF, a family of automated learning algorithms designed to optimize hyper-parameters for k local Random Forest models in large-scale classification tasks. The autokRF algorithms automatically determine an appropriate number of clusters k to partition the training dataset, then train separate Random Forest models within each cluster. This localized approach enables efficient parallel classification on multi-core CPUs. To improve model performance, each cluster applies the .632 bootstrap estimator, while hyper-parameter tuning is carried out using grid search (GS), reinforcement learning (RL), and Bayesian optimization (BO). Experiments on the ImageNet dataset (1,281,167 images, 768 features, 1,000 classes) demonstrate that our proposal (kRF, autokRF-GS, autokRF-RL, and autokRF-BO) consistently outperform the standard Random Forest algorithm, achieving significant improvements in both classification accuracy and training efficiency. On a Linux Ubuntu 24.04 system with an Intel Core i5-12400 CPU (4.4 GHz, 6 cores, 12 threads) and 32 GB of RAM, these algorithms complete the ImageNet classification task in 9.2, 10.83, 11.04, and 10.58 min respectively, with classification accuracy over 88%.