A Comparative Study of Different Dimensionality Reduction Methods with Naïve Bayes Classifier for Mapping Customer Requirements to Product Configurations

Abstract

Mapping customer requirements to product configurations are difficult due to the uncertainty and ambiguity of customers’ expression. The Naïve Bayes Classifier (NBC) is suitable to quantify the expression of customers, and to map their requirements to configurations with good performance. However, the prerequisite of manually independent of product attributes for NBC require preprocess. Dimensionality reduction methods are effective for simplifying the data complexity while separating the correlations between data Against the background, this paper conducts a comparative study of 7 dimensionality reduction methods as preprocess procedure for integrating with NBC to map customer requirements to product configurations. Two realistic design cases are illustrated for the comparison, and the outcomes are measured by the accuracy and F-measure. The results of this study imply several findings that the loss of information has great impact on all methods, and linear methods are more sensitive to the loss of information, and several nonlinear methods are more capable in handling the loss of information than other methods, and local linear methods are suggested compared with global nonlinear methods.