An Adaptive PCA-RF-WKNN Algorithm for High-Dimensional Semiconductor Defect Classification

The K-Nearest Neighbors algorithm is a widely used machine learning technique for classification and regression tasks due to its simplicity, interpretability, and effectiveness. However, KNN suffers from the curse of dimensionality and increased computational complexity when applied to high-dimensional datasets. This paper proposes an adaptive algorithm, termed PCA-RF-WKNN. It combines Principal Component Analysis and Random Forest-based feature weighting to address the challenges of KNN in high-dimensional spaces. The method reduces computational complexity by lowering the number of dimensions and improves distance calculations through weighted features. The algorithm dynamically selects the optimal number of neighbors(k)and PCA components(m)based on cross-validation. A comprehensive evaluation across four high-dimensional classification benchmark datasets—including SECOM semiconductor manufacturing, LSVT Voice Rehabilitation, Steel Plates Faults, and Leukemia Gene Expression—shows that PCA-RF-WKNN consistently outperforms both PCA-KNN and standard KNN. The statistical significance of the observed performance improvements is validated using the paired Wilcoxon signed-rank test. The primary experimental evaluation was conducted on the SECOM dataset. The results demonstrate that the proposed PCA-RF-WKNN model achieves the highest testing accuracy of 0.962, outperforming both PCA-KNN (0.923) and standard KNN (0.863). The ROC curve for PCA-RF-WKNN rises steeply toward the top-left corner, indicating strong classification capability, with an AUC of 0.900, compared to 0.816 for PCA-KNN and 0.701 for KNN. The PCA-RF-WKNN model outperforms the benchmark models across all key performance metrics, including accuracy, precision, sensitivity, F1-score, and ROC-AUC. Overall, the proposed approach significantly enhances KNN performance, surpasses PCA-KNN, and provides a promising solution for high-dimensional data problems.