Residual fault density prediction using regression methods

Abstract

Regression methods are used to model residual fault density in terms of several product and testing process measures. Process measures considered include discovered fault density, test set size and various coverage measures such as block, decision and all-uses coverage. Product measures considered include lines of code as well as block, decision and all-uses counts. The relative importance of these product/process measures for predicting residual fault density is assessed for a specific data set. Only selected testing process measures, in particular discovered fault density and decision coverage, are important predictors in this case while all product measures considered are important. These results are based on consideration of a substantial family of models, specifically, the family of quadratic response surface models with two-way interaction. Model selection is based on "leave one out at a time" cross-validation using the predicted residual sum of squares (PRESS) criterion.