Entropy-based anisotropic and isotropic regression neural network models for bug prediction
Kumari Seema Rani, Meera Sharma, Vandana Singh · Journal of Engineering Research · 2025
A code change refers to the modifications in programming constructs, namely IF Statement (IF), Method Call (MC), Method Declaration (MD), Loop (LP), Sequence (SQ), Class Field (CF), Class (C), Try (TY), Assignment (AS), Switch (SW), and Finally (F) to fix various issues (bug, new feature, and feature enhancement). The uncertainty that arises due to these changes is quantified using entropy-based measures. We computed entropy for different programming constructs using Shannon's entropy formula. We have proposed two bug prediction models, namely Model 1 and Model 2. Model 1 uses the anisotropic and isotropic Gaussian kernels of Generalized Regression Neural Network (GRNN). In Model 2, we applied an ensemble based on the weighted average of output from 11 sub-models, utilizing the entropy of each of the 11 code change categories separately. To validate the performance of the proposed models, we have considered the entropy of different programming constructs for the Core and UI components of the Eclipse project, JUnit-Framework of the JUnit project, and Megameklab, a component of the MegaMek project. Results show that Model 2 performs better than Model 1 in terms of various performance measures, namely R 2 , MSE (Mean Squared Error), MAE (Mean Absolute Error), and RMSE (Root Mean Squared Error).