An Empirical Comparison of Model-Agnostic Techniques for Defect Prediction Models

Recently, software defect prediction studies have attempted to make black-box defect prediction models explainable and actionable. State-of-the-art defect prediction studies have utilized various model-agnostic techniques derived from the explainable AI domain as key tools to make the predictions easier for practitioners to understand. However, it has not been sufficiently investigated whether there is inconsistent information within local explanations generated by different model-agnostic techniques when interpreting a defect prediction. If local explanations generated by heterogeneous model-agnostic techniques consist of different information, the derivable insights to understand and act upon defect predictions becomes less viable and it may cause ineffective or even incorrect defect corrections. In this research, we empirically analyzed 323,844 local explanations generated by three different model-agnostic techniques: (1) LIME, (2) SHAP, and (3) BreakDown. These local explanations were analyzed in terms of how the contributions of features were distributed, how the contributions were ranked, and whether the contributions were contradictory. We concluded that (i) different model-agnostic techniques provide practitioners with local explanations where average contributions of the top-ranked features are different; (ii) different model-agnostic techniques provide practitioners with local explanations consisting of different contribution rankings and inconsistent contribution directions of top-ranked features. Therefore, we recommend that practitioners should avoid using model-agnostic techniques interchangeably and must perform a multi-faceted manual validation when planning actions based on the local explanations.