Scholarlink

<p>This study investigates the impact of rule-based data preprocessing on the performance of machine learning (ML) models, focusing on recall, precision, and F1-score. In data-driven decision-making, input quality strongly influences predictive reliability, yet traditional preprocessing methods often overlook domain-specific nuances. To fill this gap, we present a new method that uses explicit rules to encode domain knowledge into the preprocessing stage. For example, we can filter out RPM ranges that are not normal or exclude motor load values that are not within acceptable limits. The proposed framework systematically refines datasets before model training, leading to improved data quality and model interpretability. Empirical evaluations show statistically significant improvements, with higher recall indicating better detection of positive cases, increased precision reflecting fewer false positives, and a balanced F1-score improvement. This research presents a scalable and reproducible framework that incorporates domain-specific rules into preprocessing, thereby improving model performance across various domains and paving the way for future enhancements, including automated rule generation and cross-domain applicability.</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

Rule-Driven Preprocessing for Improving Machine Learning Model Performance

  • 期刊论文

  • 发表日期：

    2025年12月3日

查看全文
  • Eknath Kulkarni,
  • Mayuri Digalwar

摘要

This study investigates the impact of rule-based data preprocessing on the performance of machine learning (ML) models, focusing on recall, precision, and F1-score. In data-driven decision-making, input quality strongly influences predictive reliability, yet traditional preprocessing methods often overlook domain-specific nuances. To fill this gap, we present a new method that uses explicit rules to encode domain knowledge into the preprocessing stage. For example, we can filter out RPM ranges that are not normal or exclude motor load values that are not within acceptable limits. The proposed framework systematically refines datasets before model training, leading to improved data quality and model interpretability. Empirical evaluations show statistically significant improvements, with higher recall indicating better detection of positive cases, increased precision reflecting fewer false positives, and a balanced F1-score improvement. This research presents a scalable and reproducible framework that incorporates domain-specific rules into preprocessing, thereby improving model performance across various domains and paving the way for future enhancements, including automated rule generation and cross-domain applicability.