Abstract
Machine learning has been applied successfully for faulty wafer detection tasks in semiconductor manufacturing. For the tasks, prediction models are built with prior data to predict the quality of future wafers as a function of their precedent process parameters and measurements. In real-world problems, it is common for the data to have a portion of input variables that are irrelevant to the prediction of an output variable. The inclusion of many irrelevant variables negatively affects the performance of prediction models. Typically, prediction models learned by different learning algorithms exhibit different sensitivities with regard to irrelevant variables. Algorithms with low sensitivities are preferred as a first trial for building prediction models, whereas a variable selection procedure is necessarily considered for highly sensitive algorithms. In this study, we investigate the effect of irrelevant variables on three well-known representative learning algorithms that can be applied to both classification and regression tasks: artificial neural network, decision tree (DT), and k-nearest neighbors (k-NN). We analyze the characteristics of these learning algorithms in the presence of irrelevant variables with different model complexity settings. An empirical analysis is performed using real-world datasets collected from a semiconductor manufacturer to examine how the number of irrelevant variables affects the behavior of prediction models trained with different learning algorithms and model complexity settings. The results indicate that the prediction accuracy of k-NN is highly degraded, whereas DT demonstrates the highest robustness in the presence of many irrelevant variables. In addition, a higher model complexity of learning algorithms leads to a higher sensitivity to irrelevant variables.