Application of Discriminant Analysis to the Selection of Software Cost Estimating Models

摘要

Currently, no quantitative methods exist to quantitatively select the best software cost estimating model for a particular software type or environment. By identifying the characteristics of the software that each model was best able to estimate, those characteristics could be used as a basis for predicting the best model. The analysis began by using selected models to concurrently estimate development costs for 25 known projects. Estimates from each model were compared and the most accurate model for each project was identified. The projects were assigned to the group of projects for which each model most accurately estimated development costs. After grouping each project, discriminant analysis was used to identify those input variables from all the estimating models that best discriminated between the groups. The identified input variables were then used as determinant variables as a basis to predict which model was most likely to best estimate cost for each project. The unbiased prediction rate was 60%. Despite the high prediction rate, the overall estimating accuracy was not reduced. Results indicate that use of the pre-analysis determinants to select a model would not reduce estimating error more than a random selection of models.