Using Mahalanobis–Taguchi system, logistic regression, and neural network method to evaluate purchasing audit quality

Abstract

The feature selection function in data mining facilitates the classification of vast data volumes and reduces attribute variables, enabling the construction of classification prediction models. For binary data, the Mahalanobis–Taguchi system, the logistic regression method, and the neural network method all feature high stability and accuracy. The Mahalanobis–Taguchi system differs from the other two methods in that models are developed through a measurement scale rather than from the learning of analytical data. We analyzed the audit quality of Taipei City Government public procurement from supply chains (hereafter abbreviated as public procurement) and applied the three feature selection methods to determine the items used in public procurement audit quality questionnaires. The results showed that the predictive powers of Mahalanobis–Taguchi system and logistic regression methods for the reduced question items were 93.8% and 92.5%, respectively. Furthermore, the prediction accuracy rate of the neural network method was 100%, showing that the prediction model constructed using the feature selections to reduce the number of attributes can effectively lower the number of questionnaire items while maintaining high prediction accuracy.