Scholarlink

<p>This study applies sufficient dimension reduction (SDR) techniques to develop a predictive model for classifying university students by their academic class of degree at graduation. Recognising that all course-related covariates have explanatory power for the target variable (degree classification), we utilise sliced inverse regression (SIR) to derive a few linear combinations of these covariates that retain the essential information needed for accurate prediction. A key contribution of this work is the hybridisation of the maximum entropy covariance (MEC) estimator with existing smoothed covariance estimators (SCEs), which produce a data-adaptive shrunken inverse MEC estimator to improve prediction accuracy. The estimated SIR components were used to train five classifiers in predicting Nigerian university graduates’ final classes of degrees on test data. The analysis was conducted using data from 53 graduates from the University of Ilorin, Nigeria, over a 5-year period. Results demonstrate that all classifiers based on the constructed SIR components achieved strong classification performance, with the SIR-based k-nearest neighbour (SIR-knn) classifier achieving the highest predictive accuracy of 82%.</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

Efficient multiclass prediction using sliced inverse regression with covariance hybridisation: an application to degrees classification problem

  • 期刊论文

  • 发表日期：

    2025年4月9日

查看全文
  • Kabir Opeyemi Olorede,
  • Waheed Babatunde Yahya

摘要

This study applies sufficient dimension reduction (SDR) techniques to develop a predictive model for classifying university students by their academic class of degree at graduation. Recognising that all course-related covariates have explanatory power for the target variable (degree classification), we utilise sliced inverse regression (SIR) to derive a few linear combinations of these covariates that retain the essential information needed for accurate prediction. A key contribution of this work is the hybridisation of the maximum entropy covariance (MEC) estimator with existing smoothed covariance estimators (SCEs), which produce a data-adaptive shrunken inverse MEC estimator to improve prediction accuracy. The estimated SIR components were used to train five classifiers in predicting Nigerian university graduates’ final classes of degrees on test data. The analysis was conducted using data from 53 graduates from the University of Ilorin, Nigeria, over a 5-year period. Results demonstrate that all classifiers based on the constructed SIR components achieved strong classification performance, with the SIR-based k-nearest neighbour (SIR-knn) classifier achieving the highest predictive accuracy of 82%.