Building Defect Prediction Models in Practice

  • Authors:

  • A. Gunes Koru;Hongfang Liu

  • Affiliations:

  • University of Maryland, Baltimore County;University of Maryland, Baltimore County

  • Venue:
  • IEEE Software
  • Year:
  • 2005

Quantified Score

Hi-index 0.01

Visualization

Abstract

Predicting defect-prone modules successfully can help software developers improve product quality by focusing quality assurance activities on those modules. We built several machine-learning models to predict the defective modules in five software products developed by NASA, named, CM1, JM1, KC1, KC2, and PC1. Using a set of static measures as predictor variables, the models failed to predict performance satisfactorily on the products' original data sets. However, these data sets used the smallest unit of functionality--that is, a function or method驴as a module. This meant the defect prediction was performed at a fine granularity level. Stratifying the original data sets according to module size showed the prediction performance to be better in subsets that included larger modules. Aggregating the method-level KC1 data to class level improved prediction performance for the top defect classes. Guidelines based on these results help software developers build effective defect-prediction models for focused quality assurance activities.This article is part of a special issue on predictor modeling.