Electronics, Vol. 12, Pages 1656: The Impact of Data Quality on Software Testing Effort Prediction

Electronics, Vol. 12, Pages 1656: The Impact of Data Quality on Software Testing Effort Prediction

Electronics doi: 10.3390/electronics12071656

Authors: Łukasz Radliński

Background: This paper investigates the impact of data quality on the performance of models predicting effort on software testing. Data quality was reflected by training data filtering strategies (data variants) covering combinations of Data Quality Rating, UFP Rating, and a threshold of valid cases. Methods: The experiment used the ISBSG dataset and 16 machine learning models. A process of three-fold cross-validation repeated 20 times was used to train and evaluate each model with each data variant. Model performance was assessed using absolute errors of prediction. A ‘win–tie–loss’ procedure, based on the Wilcoxon signed-rank test, was applied to identify the best models and data variants. Results: Most models, especially the most accurate, performed the best on a complete dataset, even though it contained cases with low data ratings. The detailed results include the rankings of the following: (1) models for particular data variants, (2) data variants for particular models, and (3) the best-performing combinations of models and data variants. Conclusions: Arbitrary and restrictive data selection to only projects with Data Quality Rating and UFP Rating of ‘A’ or ‘B’, commonly used in the literature, does not seem justified. It is recommended not to exclude cases with low data ratings to achieve better accuracy of most predictive models for testing effort prediction.