<p>High-dimensional multi-label datasets often exhibit complex and nonlinear dependencies among features and labels, posing significant challenges for effective feature selection (FS). While filter-based approaches offer scalability, they often overlook label structure, whereas traditional wrapper methods incur high computational costs due to repeated classifier training. To address these limitations, we propose a novel final-model-independent wrapper-based FS method, grey wolf optimizer with adaptive operator selection for multi-label feature selection (GWO-AOSMFS). The method employs a binary grey wolf optimizer enhanced with a reinforcement-based adaptive mutation strategy to effectively explore the discrete feature space. A multi-criteria fitness function is formulated to jointly model linear feature–label correlations, nonlinear relevance via entropy-weighted mutual information, and label co-occurrence preservation through cosine similarity using a lightweight kNN projection mechanism. Unlike traditional wrapper methods that require repeated training of complex classifiers, GWO-AOSMFS evaluates feature subsets using a composite fitness function that avoids training any final classifier. A lightweight, training-free kNN module is used only to project labels for co-occurrence preservation, not for predictive modeling. To promote compactness and interpretability, a sparsity-inducing penalty is included. This design enables robust evaluation of feature subsets without the computational burden of iterative classifier invocation. Extensive experiments on thirteen benchmark datasets demonstrate that GWO-AOSMFS achieves the highest <i>Acc</i> 0.5155, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(Micro-F_1\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>M</mi> <mi>i</mi> <mi>c</mi> <mi>r</mi> <mi>o</mi> <mo>-</mo> <msub> <mi>F</mi> <mn>1</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> 0.5748, and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(Macro-F_1\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>M</mi> <mi>a</mi> <mi>c</mi> <mi>r</mi> <mi>o</mi> <mo>-</mo> <msub> <mi>F</mi> <mn>1</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> 0.5209, along with the lowest <i>HL</i> 0.0287, outperforming nine state-of-the-art methods. Statistical significance testing and stability analysis further affirm its robustness and generalizability across diverse multi-label datasets.</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-label feature selection via entropy-aware correlation metrics and adaptive operator-guided grey wolf optimization

  • Gurudatta Verma,
  • Tirath Prasad Sahu

摘要

High-dimensional multi-label datasets often exhibit complex and nonlinear dependencies among features and labels, posing significant challenges for effective feature selection (FS). While filter-based approaches offer scalability, they often overlook label structure, whereas traditional wrapper methods incur high computational costs due to repeated classifier training. To address these limitations, we propose a novel final-model-independent wrapper-based FS method, grey wolf optimizer with adaptive operator selection for multi-label feature selection (GWO-AOSMFS). The method employs a binary grey wolf optimizer enhanced with a reinforcement-based adaptive mutation strategy to effectively explore the discrete feature space. A multi-criteria fitness function is formulated to jointly model linear feature–label correlations, nonlinear relevance via entropy-weighted mutual information, and label co-occurrence preservation through cosine similarity using a lightweight kNN projection mechanism. Unlike traditional wrapper methods that require repeated training of complex classifiers, GWO-AOSMFS evaluates feature subsets using a composite fitness function that avoids training any final classifier. A lightweight, training-free kNN module is used only to project labels for co-occurrence preservation, not for predictive modeling. To promote compactness and interpretability, a sparsity-inducing penalty is included. This design enables robust evaluation of feature subsets without the computational burden of iterative classifier invocation. Extensive experiments on thirteen benchmark datasets demonstrate that GWO-AOSMFS achieves the highest Acc 0.5155, \(Micro-F_1\) M i c r o - F 1 0.5748, and \(Macro-F_1\) M a c r o - F 1 0.5209, along with the lowest HL 0.0287, outperforming nine state-of-the-art methods. Statistical significance testing and stability analysis further affirm its robustness and generalizability across diverse multi-label datasets.