A Reproducible, Imbalance-Aware and Explainable Software Defect Prediction Framework

Software defect prediction can assist software testing teams in determining which modules require targeting to estimate the failures accurately before any defect detection in the software.This implementation paper provides information about the aspects of defect-prediction research (provided in each of the sections) through the design of an implementation workflow.The implementation consists of data extraction, leakage-safe pre-processing, project-aware split, class imbalance, feature selection, heterogeneous ML models, threshold selection, multi-metrics evaluation, explanation, and robustness testing.The implementation is intentionally modularized; the researcher can start from the tabular Random Forest or logistic-regression baseline and later on add XGBoost, multilayer perceptron, call-graph embeddings or program-graph features without the need to change the evaluation process.The end-to-end implementation consists of 5-stages in the workflow including data collection, preprocessing, model building, performance evaluation and inference.It also presents an experiment matrix for the study of cross-project generalization, imbalance strategies, feature selection, and ensembles, as well as decision thresholds, explainability, and metamorphic validation.Since the provided archive consists of research papers and not a shared standardized dataset and execution environment, the numerical results from the literature are treated as reference evidence while not being newly generated benchmark results.Thus, the implementation emphasizes reproducibility and a distinct separation between what we report as evidence and experimentation that can be executed.