Comparison of M5’ model tree with MLR in the development of fault prediction models involving interaction between metrics

Abstract

Amongst the critical actions needed to be undertaken before system testing, software fault prediction is imperative. Prediction models are used to identify fault-prone classes and contribute considerably to reduce the testing time, project risks, and resource and infrastructure costs. In the development of a prediction model, the interaction of metrics results in an improved predictive capability, accruing to the fact that metrics are often correlated and do not have a strict additive effect in a regression model. Even though the interaction amongst metrics results in the model’s improved prediction capability, it also gives rise to a large number of predictors. This leads to Multiple Linear Regression (MLR) exhibiting a reduced level of performance, since a single predictive formula occupies the entire data space. The M5’ model tree has an edge over MLR in managing such interactions, by partitioning the data space into smaller regions. The resulting hypothesis empirically establish that the M5’ model tree, when applied to these interactions, provides a greater degree of accuracy and robustness of the model as a whole when compared with MLR models.