Scholarlink

<p>Alcohol stability is a key safety indicator for the thermal processing adaptability of dairy products. Traditional alcohol testing is cost-effective, but its time and labor costs accumulate significantly given the huge daily intake of raw milk. To reduce detection costs, this study proposes a machine learning-based method using easily accessible physicochemical indicators. After evaluating seven mainstream models, RF, MLP, and ReconTab were identified as the most promising base learners. Then, a novel ensemble model, RMRT_embedding, which integrates these models via stacking with self-supervised ReconTab embeddings, was developed to enhance generalization under imbalanced data. Although the indicators of positive samples may benefit from the limitations of the sampling strategy, however, experiments on the full dataset show that our method demonstrates extremely high performance when dealing with a large number of unseen negative samples, thereby having a better capacity for suppressing false positives. The results confirmed that our method is effective and explainable, thus providing an economical and effective tool for the quality control of dairy products.</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

Leveraging Machine Learning and Model Stacking for Robust Prediction of Alcohol Stability in Raw Milk

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • Haohan Ding,
  • Huadi Huang,
  • Xiaodong Song,
  • Xiaohui Cui,
  • David I. Wilson,
  • Wei Yu,
  • Song Shen,
  • Zhiran Liang,
  • Guanjun Dong

摘要

Alcohol stability is a key safety indicator for the thermal processing adaptability of dairy products. Traditional alcohol testing is cost-effective, but its time and labor costs accumulate significantly given the huge daily intake of raw milk. To reduce detection costs, this study proposes a machine learning-based method using easily accessible physicochemical indicators. After evaluating seven mainstream models, RF, MLP, and ReconTab were identified as the most promising base learners. Then, a novel ensemble model, RMRT_embedding, which integrates these models via stacking with self-supervised ReconTab embeddings, was developed to enhance generalization under imbalanced data. Although the indicators of positive samples may benefit from the limitations of the sampling strategy, however, experiments on the full dataset show that our method demonstrates extremely high performance when dealing with a large number of unseen negative samples, thereby having a better capacity for suppressing false positives. The results confirmed that our method is effective and explainable, thus providing an economical and effective tool for the quality control of dairy products.