Optimal Feature Subset Determination for High-Dimensional Datasets in Manufacturing Processes

Selecting the most relevant feature subset with good predictive performances and computational speed for high-dimensional datasets is usually challenging in manufacturing processes for product quality control. In this paper, an optimal feature subset determination approach for high-dimensional datasets is proposed. The algorithm starts with ranking the importance of individual features using the normalized mutual information concepts. Then, through optimizing a new evaluation metric, an initial relevant feature subset is obtained. Finally, redundant features are eliminated to help obtain the final selection results. The proposed method was successfully used in solving a real-world high-dimensional feature selection problem in the semiconductor industry. Comparisons were made with four other representative feature selection algorithms, including Relief-F, mRMR, FCBF, and IWFS in processing a number of datasets from different applications. It is demonstrated that the proposed method can automatically determine an optimal feature subset to achieve good average predictive accuracy with less computational resources.