Scholarlink

<p>Data preprocessing is a pivotal stage in Machine Learning (ML) workflows, directly influencing both predictive performance and computational efficiency. Although widely acknowledged, the quantitative impact of preprocessing choices, particularly dimensionality reduction strategies, remains insufficiently explored in classification tasks. This study compares three data preparation approaches: raw data, Principal Component Analysis (PCA), and Principal Component Factor Analysis (PCFA). While PCA and Factor Analysis (FA) are well-established techniques, their structured integration as PCFA has received limited attention. By evaluating these alternatives, this study quantifies how dimensionality reduction affects predictive capability, result stability, and processing time across diverse learning scenarios. A factorial Design of Experiments (DOE) framework is employed, enabling the assessment of model performance under multiple attribute configurations and their interaction effects. Ten distinct ML algorithms are applied to ten heterogeneous datasets from different domains, with model training performed using cross-validation and hyperparameter optimization via Grid Search and Optuna. The findings show that PCA and PCFA consistently achieve competitive predictive performance while substantially reducing computational cost compared to raw data. Furthermore, data structure and attribute interactions strongly influence outcomes, indicating that the relationship between preprocessing strategies and model behavior is scenario-dependent. We conclude that combining dimensionality reduction, structured feature design, and optimized training strategies is essential for balancing predictive accuracy and computational cost, a trade-off increasingly critical in modern ML applications.</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

Machine learning optimization: how preprocessing strategies shape predictive performance and computational efficiency

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月3日

查看全文
  • Matheus Costa Pereira,
  • Mirelli de Castro Cesário,
  • Vinicius Paes de Carvalho,
  • Matheus Brendon Francisco,
  • João Luiz Junho Pereira,
  • Anderson Paulo de Paiva

摘要

Data preprocessing is a pivotal stage in Machine Learning (ML) workflows, directly influencing both predictive performance and computational efficiency. Although widely acknowledged, the quantitative impact of preprocessing choices, particularly dimensionality reduction strategies, remains insufficiently explored in classification tasks. This study compares three data preparation approaches: raw data, Principal Component Analysis (PCA), and Principal Component Factor Analysis (PCFA). While PCA and Factor Analysis (FA) are well-established techniques, their structured integration as PCFA has received limited attention. By evaluating these alternatives, this study quantifies how dimensionality reduction affects predictive capability, result stability, and processing time across diverse learning scenarios. A factorial Design of Experiments (DOE) framework is employed, enabling the assessment of model performance under multiple attribute configurations and their interaction effects. Ten distinct ML algorithms are applied to ten heterogeneous datasets from different domains, with model training performed using cross-validation and hyperparameter optimization via Grid Search and Optuna. The findings show that PCA and PCFA consistently achieve competitive predictive performance while substantially reducing computational cost compared to raw data. Furthermore, data structure and attribute interactions strongly influence outcomes, indicating that the relationship between preprocessing strategies and model behavior is scenario-dependent. We conclude that combining dimensionality reduction, structured feature design, and optimized training strategies is essential for balancing predictive accuracy and computational cost, a trade-off increasingly critical in modern ML applications.