یادگیری ماشین با حساسیت هزینه برای پیش‌بینی نقص‌های تولید: رویکردی نوین مبتنی بر metacost

کنترل کیفیت و کاهش هزینه‌های تولید، به پیش‌بینی دقیق عیوب در فرآیندهای صنعتی وابسته است. در این پژوهش، رویکرد یادگیری ماشین حساس به هزینه، با استفاده از الگوریتم metacost بررسی شده است. metacost یک تکنیک پس‌پردازش برای تبدیل مدل‌های یادگیری ماشین به مدل‌های حساس به هزینه است که با در نظر گرفتن ماتریس هزینۀ خطاها، تصمیم‌گیری مدل را بهینه می‌کند. هدف اصلی، کاهش خطاهای منفی کاذب در شناسایی روزهای پرنقص تولید است. برای این منظور، از چندین الگوریتم شامل random forest، gradient boosting، xgboost، lightgbm، catboost، svm و رگرسیون لجستیک استفاده شد. داده‌ها از دیتاست «predicting manufacturing defects» برگرفته از پلتفرم kaggle، شامل اطلاعات مربوط به 3240 روز تولید صنعتی جمع‌آوری شدند. نتایج نشان داد که الگوریتم جنگل تصادفی با دستیابی به صحت برابر 96.9% و بازخوانی برابر 98.9‎‍%، بهترین عملکرد را در میان مدل‌ها داشت. به‌ویژه‎‍ توانایی بالای این مدل در شناسایی صحیح روزهای پرنقص، آن را به گزینۀ مناسبی برای کاربردهای واقعی در صنعت تبدیل کرد. دیگر مدل‌ها نیز عملکرد پذیرفتنی داشتند؛ اما در مقایسه با random forest، در کاهش نرخ منفی کاذب ضعیف‌تر ظاهر شدند. این نتایج، کارایی رویکردهای حساس به هزینه را در بهبود پیش‌بینی نقص تولید، تایید می‌کند.

cost-sensitive machine learning for predicting production defects: a novel approach based on metacost

purpose: this paper aims to apply cost-sensitive machine learning (csml) for predicting production defects in industrial processes to minimize the costs associated with false negatives. it investigates the metacost algorithm to assess its effectiveness in enhancing accuracy for defect detection when misclassification costs are incorporated into predictive models.design/methodology/approach: real-world industrial production data is utilized to train multiple machine learning models: random forest, xgboost, lightgbm, catboost, gradient boosting, svm, and logistic regression to predict high-defect days. subsequently, we implement various techniques to address class imbalance and enhance generalization, including borderline-smote, recursive feature elimination with cross validation (rfecv), and optuna hyperparameter optimization. after modeling, metacost is applied as a post-processing step to convert base learners into cost-sensitive classifiers with a 5:1 cost ratio focused on minimizing false negatives.findings: findings indicated that among all tested classifiers, the most effective model for identifying high defect production days is the random forest model combined with metacost, as it boasts the highest recall (98.9%). in industrial settings, this high recall is particularly crucial, as failing to detect defects can lead to significant losses, waste, and quality degradation for the company. catboost and lightgbm also performed well in terms of precision and false negatives, matching random forest’s performance, but they recorded slightly more false negatives. additionally, the random forest model achieved a balanced f1 score, making it reliable and robust for imbalanced datasets. this highlights the importance of cost-sensitive models in quality control tasks, especially when the cost of false negatives far exceeds that of false positives. random forest consistently outperforms in this study, making it a promising candidate for practical application in real-world manufacturing, as it can predict problematic production days, prompting timely interventions and continuous process improvement. the findings thus confirm both the strategic and technical value of the model for industrial decision-making processes.research limitations/implications: the industrial variations in the real world may not be fully represented by the study’s dataset. results are confined to batch-based manufacturing data and require validation in continuous or real-time processes. future research should concentrate on deep learning models (cnn or rnn) and real-time sensor data or iot-based streams for dynamic settings. practical implications: this study offers a data-driven framework for manufacturers to predict and address production defects, thereby reducing waste and rework costs. the approach minimizes undetected high-defect days while enhancing production quality, decreasing operational losses, and enabling the use of simple ml-based solutions instead of costly diagnostic equipment. social implications: enhanced defect prediction for sustainable manufacturing is achieved through reduced material waste, prolonged product lifespan, and improved quality control. greater product reliability fosters stronger consumer trust and diminishes the environmental impact of industrial waste.originality/value: the metacost initiative has implemented this algorithm as one of the first of its kind in the real industrial defect prediction domain. cost-sensitive learning formulations with modern ml tools are employed to bridge a gap in research that addresses production optimization while balancing quality and cost. this research is valuable both academically and for industry practitioners interested in economic considerations in predictive modeling. furthermore, while previous works focused primarily on healthcare, finance, and software defect detection, this research introduces a novel application of metacost in a manufacturing environment characterized by highly imbalanced class distributions and complex processes. the study proposes a replicable and scalable pipeline by integrating techniques such as borderline-smote, rfecv, and optuna, which can be utilized for future research as well as in real-time applications. it effectively closes the gap between theoretical ml design and the application of ml in high-impact quality improvement of industrial systems.