<p>Stability, a method’s sensitivity to training set changes, is often overlooked in studies, despite its importance in robustness and model performance. Our proposal aims to address this issue without compromising feature reduction. The proposed model uses an NSGA-II multiobjective genetic algorithm to optimise two competing goals, accuracy and number of features. The suggested model functions in two phases. The first is stability retainment, population initialization and, multiobjective optimization being the second stage. A hybrid ensemble technique that integrates both data and functional variation is applied to select stable features. Functional variation is achieved using three different feature selection algorithms: mRMR, ReliefF, and Random forest. The selected stable feature subsets guide the population initialization process of the multi-objective NSGA-II algorithm to optimize two objectives, accuracy and number of features. We considered three datasets, two being gene expressions in a microarray dataset, Leukemia and colon cancer, and one is an imbalanced dataset, Cervical cancer risk factor to validate the proposed method. The results show a significant increase in accuracy and reduced feature subset size with the same or improved stability.</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

RBON: A Robust Bi-objective Feature Selection Framework Using Hybrid Ensemble Technique and NSGA-II

  • Kurman Sangeeta,
  • Sumitra Kisan

摘要

Stability, a method’s sensitivity to training set changes, is often overlooked in studies, despite its importance in robustness and model performance. Our proposal aims to address this issue without compromising feature reduction. The proposed model uses an NSGA-II multiobjective genetic algorithm to optimise two competing goals, accuracy and number of features. The suggested model functions in two phases. The first is stability retainment, population initialization and, multiobjective optimization being the second stage. A hybrid ensemble technique that integrates both data and functional variation is applied to select stable features. Functional variation is achieved using three different feature selection algorithms: mRMR, ReliefF, and Random forest. The selected stable feature subsets guide the population initialization process of the multi-objective NSGA-II algorithm to optimize two objectives, accuracy and number of features. We considered three datasets, two being gene expressions in a microarray dataset, Leukemia and colon cancer, and one is an imbalanced dataset, Cervical cancer risk factor to validate the proposed method. The results show a significant increase in accuracy and reduced feature subset size with the same or improved stability.