Scholarlink

<p>Kernel Ridge Regression (E-KRR) is a widely used method for modeling nonlinear relationships, but it often suffers from manual kernel selection and sensitivity to parameter settings. These issues can lead to poor generalization and unstable performance across different datasets. In this paper, we propose an improved method called Ensemble Kernel Ridge Regression (E-KRR), which addresses these limitations by incorporating a multi-view learning approach. We begin by modeling data as having multiple views and extend traditional ridge regression to this setting. These views are then mapped into multiple kernel representations within different Reproducing Kernel Hilbert Spaces (RKHSs). E-KRR automatically learns optimal combinations of kernels and their weights directly from data, avoiding manual kernel tuning. Experimental results on 16 datasets demonstrate that E-KRR consistently outperforms several state-of-the-art methods. It achieves up to 11.4% lower MSE in regression tasks and improves classification accuracy by 2.8–4.3% on image datasets and 1.1–9.3% on tabular datasets, confirming its robustness and effectiveness.</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-view data-driven ensemble kernel ridge regression via multi-kernel optimization

  • 期刊论文

  • 发表日期：

    2026年7月21日

查看全文
  • Kun Qu,
  • Zhi-Feng Liu,
  • Emmanuel Ntaye,
  • Ernest Domanaanmwi Ganaa,
  • Xiang-Jun Shen

摘要

Kernel Ridge Regression (E-KRR) is a widely used method for modeling nonlinear relationships, but it often suffers from manual kernel selection and sensitivity to parameter settings. These issues can lead to poor generalization and unstable performance across different datasets. In this paper, we propose an improved method called Ensemble Kernel Ridge Regression (E-KRR), which addresses these limitations by incorporating a multi-view learning approach. We begin by modeling data as having multiple views and extend traditional ridge regression to this setting. These views are then mapped into multiple kernel representations within different Reproducing Kernel Hilbert Spaces (RKHSs). E-KRR automatically learns optimal combinations of kernels and their weights directly from data, avoiding manual kernel tuning. Experimental results on 16 datasets demonstrate that E-KRR consistently outperforms several state-of-the-art methods. It achieves up to 11.4% lower MSE in regression tasks and improves classification accuracy by 2.8–4.3% on image datasets and 1.1–9.3% on tabular datasets, confirming its robustness and effectiveness.