Machine Learning to Evaluate Evolvability Defects: Code Metrics Thresholds for a Given Context

摘要

Evolvability defects are non-understandable and non-modifiable states that do not directly produce runtime behavioral failures. Automatic source code evaluation by metrics and thresholds can help reduce the burden of a manual inspection. This study addresses two problems. (1) Evolvability defects are not usually managed in bug tracking systems. (2) Conventional methods cannot fully interpret the relations among the metrics in a given context (e.g., programming language, application domain). The key actions of our method are to (1) gather training-data for machine learning by experts' manual inspection of some of the files in given systems (benchmark) and (2) employ a classification-tree learner algorithm, C5.0, which can deal with non-orthogonal relations between metrics. Furthermore, we experimentally confirm that, even with less training-data, our method provides a more precise evaluation than four conventional methods (the percentile, Alves' method, Bender's method, and the ROC curve-based method).