Software defect prediction (SDP) plays a pivotal role by making sure that software is reliable and of high quality. It does this by identifying parts of software that are likely to make mistakes early in the development phase. Traditional machine learning models, on the other hand, often don’t work well with real-world SDP datasets because they have an imbalance in classes. This study proposes an explainable and lightweight ensemble learning approach that uses integration of advanced resampling and feature selection methods to address the problem of imbalanced classification. The method uses both Mutual Information (MI) and Recursive Feature Elimination (RFE) feature selection, which successfully reduce noise and dimensionality. The main prediction model uses a group of soft-voting classifiers like (Logistic Regression, Decision Tree, and Gaussian Naive Bayes). It also uses AdaBoost and Balanced Random Forest (BRF) to enhance robustness. Experimental evaluation across 24 datasets from the PROMISE and NASA repositories and found that it gives good results in terms of AUC, F1-score, and G-mean. Also, local interpretability is achieved with LIME, which provides model transparency and insights that can be used. The framework’s efficiency, interpretability, and adaptability make it useful for practical defect prediction in software engineering.