An Evaluation of Parameter Pruning Approaches for Software Estimation

Abstract

Model-based estimation often uses impact factors and historical data to predict the effort of new projects. Estimation accuracy of this approach is highly dependent on how well impact factors are selected. This paper comparatively assesses six methods for prune parameters of effort estimation models, including Stepwise regression, Lasso, constrained regression, GRASP, Tabu search, and PCA. Four data sets were used for evaluation, showing that estimation accuracy varies among the methods but no method consistently outperforms the rest. Stepwise regression prunes estimation model parameters the most while it does not sacrifice much estimation performance. Our study provides further evidence to support the use of Stepwise regression for selecting factors in effort estimation.