<![CDATA[The semiconductor manufacturing industry faces challenges in efficient and accurate product quality control. Traditional manual inspection methods are still used but have limitations in speed and accuracy. Therefore, the application of machine learning has become a potential solution to improve the prediction of semiconductor production outcomes. This study develops a predictive model using Logistic Regression, Decision Tree, Random Forest, XGBoost, Naïve Bayes, and Support Vector Machine (SVM) with various dataset partitioning techniques such as Normal Data (90:10), Oversampling (70:30), Undersampling (80:20 & 70:30), and Principal Component Analysis (PCA) (90:10). The dataset used is sourced from the Factory Manufacturing Semiconductor Test (FMST), comprising 1,567 samples and 591 features, with product quality test labels (Pass/Fail). The results show that XGBoost and Random Forest achieved the highest accuracy (0.95) on the normal dataset (90:10), while Naïve Bayes had the lowest performance (0.23) due to its limitations in handling datasets with a large number of features. The oversampling technique improved the performance of Decision Tree and Logistic Regression but reduced the accuracy of XGBoost due to the risk of overfitting. Meanwhile, undersampling was more effective for Decision Tree but decreased SVM performance. The application of PCA improved Logistic Regression accuracy to 0.81, proving that dimensionality reduction can enhance model efficiency. Further analysis shows that feature selection using F-score can optimize model performance by eliminating redundant features. This study concludes that XGBoost and Random Forest are the best models for predicting semiconductor manufacturing product outcomes, with broad potential applications in the industry.]]>
Copyrights © 2025
