An empirical study of the impact of count models predictions on module-order models

摘要

Software quality prediction models are used to achieve high software reliability. A module-order model (MOM) uses an underlying quantitative prediction model to predict this rank-order. This paper compares performances of module-order models of two different count models which are used as the underlying prediction models. They are the Poisson regression model and the zero-inflated Poisson regression model. It is demonstrated that improving a count model for prediction does not ensure a better MOM performance. A case study of a full-scale industrial software system is used to compare performances of module-order models of the two count models. It was observed that improving prediction of the Poisson count model by using zero-inflated Poisson regression did not yield module-order models with better performance. Thus, it was concluded that the degree of prediction accuracy of the underlying model did not influence the results of the subsequent module-order model. Module-order modeling is proven to be a robust and effective method even though both underlying prediction may sometimes lack acceptable prediction accuracy.