고차원 희소 반도체 데이터 이상 탐지를 위한 차원 축소 및 불균형 처리 기법 비교

초록

Defect detection in semiconductor manufacturing faces significant challenges due to high-dimensional, sparse, and imbalanced process data. This study systematically evaluates the performance of three dimensionality reduction techniques (PCA, TruncatedSVD, and correlation-based feature selection) combined with two imbalance handling approaches(SMOTE-NC and algorithm-level class weighting) across four classifiers. Experiments were conducted on manufacturing data characterized by 1,558 features, 98.99% sparsity, and a 1:11.33 class imbalance ratio. Results demonstrate that the combination of correlation-based feature selection and algorithm-level class weighting achieved the highest performance. Repeated experiments further confirmed that SVM and LR are statistically equivalent top- performing models within this pipeline (p=0.1289), with SVM recording a G-Mean of 0.7843 in the baseline experiment. The findings indicate that target-oriented feature selection outperforms variance-based methods in sparse environments, and direct loss function adjustment proves more robust than synthetic oversampling for highly compressed feature spaces. This research provides empirical guidelines for optimizing preprocessing strategies in industrial defect prediction.

키워드