ASFE Feature Selection Technique Integrated with SMOTE-ENN for Handling High-Dimensional and Highly Imbalanced Wafer Defect Classification Data

In semiconductor manufacturing, the advancement of high-throughput sensors leads to the generation of high-dimensional and highly imbalanced data, presenting challenges like feature redundancy, increased computational costs, and bias in the classification models towards the majority class. These challenges complicate wafer defect classification, necessitating the integration of advanced feature selection and sampling techniques. Therefore, this paper introduces the Adaptive Simple, Fast, and Efficient (ASFE) model, an adaptive evolutionary feature selection technique integrated with SMOTE-ENN resampling. This model enhances the Simple, Fast, Efficient (SFE) feature selection model to address semiconductor data's high-dimensional and highly imbalanced challenges. The ASFE model, integrated with SMOTE-ENN resampling, is benchmarked against baseline models such as SFE and SMOTE-ENN alone, as well as advanced evolutionary feature selection algorithms like Binary Particle Swarm Optimization (BPSO) and Binary Differential Evolution (BDE). The experiment results based on a semiconductor's wafer defect classification dataset demonstrate that ASFE achieves the highest G-mean and F -measure scores with reduced computational time. Although its AUC score is slightly lower than that of SMOTE-ENN, the variability indicated by its standard deviation suggests that ASFE could outperform under certain conditions. These findings underscore the capability of integrating the ASFE model with SMOTE-ENN resampling to enhance the accuracy and efficiency of wafer defect classification in semiconductor manufacturing.