Neural Net Analysis of the Propensity for Change in Large Software Systems

A novel approach for analyzing the relationship between code metrics and change count histories is presented. Specifically, neural networks are employed to determine a mapping between metrics and change count. While these neural networks can be trained to a high degree of accuracy, their internal workings remain opaque to the user. As such, a fuzzy modeling approach is additionally employed to generate the rules governing the neural computation. These rules are linguistic in nature and are thus more easily interpreted by software project managers. Application of this method to Mozilla change data reveals the importance of fan-out, total lines of code and maximum cyclomatic complexity metrics in predicting amount of change per file.