Classification-tree models of software-quality over multiplereleases

摘要

This paper presents an empirical study that evaluatesnsoftware-quality models over several releases, to address the question,n“How long will a model yield useful predictions?” Thenclassification and regression trees (CART) algorithm is introduced, CARTncan achieve a preferred balance between the two types ofnmisclassification rates. This is desirable because misclassification ofnfault-prone modules often has much more severe consequences thannmisclassification of those that are not fault-prone. The case-studyndeveloped 2 classification-tree models based on 4 consecutive releasesnof a very large legacy telecommunication system. Forty-two softwarenproduct, process and execution metrics were candidate predictors. Modeln1 used measurements of the first release as the training data set; thisnmodel had 11 important predictors. Model 2 used measurements of thensecond release as the training data set; this model had 15 importantnpredictors. Measurements of subsequent releases were evaluation datansets. Analysis of the models' predictors yielded insights into variousnsoftware development practices. Both models had accuracy that would benuseful to developers. One might suppose that software-quality modelsnlose their value very quickly over successive releases due to evolutionnof the product and the underlying development processes. The authorsnfound the models remained useful over all the releases studied