Scholarlink

<p>Feature Selection (FS) is a crucial preprocessing step that enhances model performance by eliminating redundant and irrelevant features. However, high-dimensional data introduces the "curse of dimensionality," characterized by an exponentially expanding search space, data sparsity, model overfitting, and prohibitive computational costs. While the Northern Goshawk Optimization (NGO) algorithm is promising due to its simplicity and minimal parameters, its efficacy in FS is limited by insufficient search diversity and a propensity for local optimum in high-dimensional spaces. To address these challenges, this paper proposes a binary two-phase strategy-enhanced NGO (bTENGO) algorithm. Our approach systematically mitigates the curse of dimensionality through a dual-phase mechanism: (1) Global Exploration: A hybrid sine–cosine strategy with randomized parameters is introduced to diversify search directions and prevent premature convergence. This is coupled with a guided position update strategy that dynamically adjusts step sizes toward global optimum, significantly boosting search efficiency. (2) Local exploitation: a nonlinear dual exploitation strategy, integrated with elite-based neighborhood search, balances exploration and exploitation. This enables rapid, stable convergence to high-quality solutions with refined precision. Comprehensive experiments on 14 high-dimensional datasets from the UCI repository demonstrate that bTENGO achieves superior feature reduction while maintaining competitive classification accuracy. Comparative results against nine state-of-the-art FS algorithms confirm that bTENGO significantly alleviates the challenges of high-dimensionality and delivers robust, high-performance feature selection.</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

Two-phase strategy-enhanced northern goshawk optimization algorithm for high-dimensional feature selection

  • 期刊论文

  • 发表日期：

    2025年11月15日

查看全文
  • Xinlu Zong,
  • Fucai Liu,
  • Jiaxin Hao

摘要

Feature Selection (FS) is a crucial preprocessing step that enhances model performance by eliminating redundant and irrelevant features. However, high-dimensional data introduces the "curse of dimensionality," characterized by an exponentially expanding search space, data sparsity, model overfitting, and prohibitive computational costs. While the Northern Goshawk Optimization (NGO) algorithm is promising due to its simplicity and minimal parameters, its efficacy in FS is limited by insufficient search diversity and a propensity for local optimum in high-dimensional spaces. To address these challenges, this paper proposes a binary two-phase strategy-enhanced NGO (bTENGO) algorithm. Our approach systematically mitigates the curse of dimensionality through a dual-phase mechanism: (1) Global Exploration: A hybrid sine–cosine strategy with randomized parameters is introduced to diversify search directions and prevent premature convergence. This is coupled with a guided position update strategy that dynamically adjusts step sizes toward global optimum, significantly boosting search efficiency. (2) Local exploitation: a nonlinear dual exploitation strategy, integrated with elite-based neighborhood search, balances exploration and exploitation. This enables rapid, stable convergence to high-quality solutions with refined precision. Comprehensive experiments on 14 high-dimensional datasets from the UCI repository demonstrate that bTENGO achieves superior feature reduction while maintaining competitive classification accuracy. Comparative results against nine state-of-the-art FS algorithms confirm that bTENGO significantly alleviates the challenges of high-dimensionality and delivers robust, high-performance feature selection.