Scholarlink

<p>This article proposes a new likelihood ratio-based index (LR index for short), to measure the dependence between a categorical response variable and a continuous predictor variable. The LR index is nonnegative and is zero if and only if the variables are independent. We propose an estimate of the index, develop a novel independence test and derive the asymptotic null distribution. Next, based on the LR index, a feature screening procedure (LR-SIS for short) is developed for multiclass classification with ultrahigh-dimensional predictors. LR-SIS is model-free and robust to the heavy-tailed distribution of predictors and outliers. The sure screening property of LR-SIS is established allowing the number of response classes to be diverging. The finite sample performance of the proposed LR index in both independence testing and feature screening is demonstrated by comprehensive simulation studies. Application of the LR-SIS is also illustrated on a real data set.</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

Model-Free Feature Screening for Ultrahigh-Dimensional Multiclass Classification via a Likelihood Ratio-Based Measure of Dependence

  • 期刊论文

  • 发表日期：

    2026年3月28日

查看全文
  • Fei Ye,
  • Weidong Ma,
  • Jingsong Xiao,
  • Ying Yang

摘要

This article proposes a new likelihood ratio-based index (LR index for short), to measure the dependence between a categorical response variable and a continuous predictor variable. The LR index is nonnegative and is zero if and only if the variables are independent. We propose an estimate of the index, develop a novel independence test and derive the asymptotic null distribution. Next, based on the LR index, a feature screening procedure (LR-SIS for short) is developed for multiclass classification with ultrahigh-dimensional predictors. LR-SIS is model-free and robust to the heavy-tailed distribution of predictors and outliers. The sure screening property of LR-SIS is established allowing the number of response classes to be diverging. The finite sample performance of the proposed LR index in both independence testing and feature screening is demonstrated by comprehensive simulation studies. Application of the LR-SIS is also illustrated on a real data set.