Comparative Evaluation of TabKANet with Oversampling and Feature Selection Ablation for Software Defect Prediction

TabKANet for Software Defect Prediction: An Oversampling and Feature-Selection Ablation Study

Optimizing Pi–Sigma Neural Networks for software defect prediction via correlation-based feature selection and robust evaluation

Software defect prediction is essential for maintaining code quality in critical domains, yet it remains challenging due to feature redundancy and class imbalance. This study proposes an optimized Pi–Sigma Neural Network (PSNN) framework leveraging Correlation-Based Feature Selection (CBFS) and Min-Max normalization. Utilizing the NASA PROMISE CM1 dataset, a 5-fold stratified cross-validation pipeline was implemented to ensure statistical robustness and prevent data leakage. Experimental results on the CM1 dataset show the refined PSNN achieves high performance (99.80% accuracy on CM1) after aggressive feature reduction to 3–5 features and a precision of 1.000. To address class imbalance, the model achieved a Matthews Correlation Coefficient (MCC) of 0.988 and a G-Mean of 0.990. Comparative analysis shows that the developed PSNN-FS, despite its simplicity, achieves strong performance competitive with more complex architectures on the CM1 dataset.

COMPARATIVE EVALUATION OF CLASS-IMBALANCE CORRECTION TECHNIQUES FOR SOFTWARE DEFECT PREDICTION

Class imbalance can make software defect predictors appear successful while missing defective modules. This study compared no correction, random oversampling, random undersampling, SMOTE, ADASYN, and class-weighted learning across logistic regression, decision tree, random forest, support vector machine, and neural network classifiers. KC1 and PC1 NASA/PROMISE datasets (3,218 modules; 403 defective) were evaluated by stratified five-fold cross-validation. Imputation, scaling, and correction were fitted only within training folds. Precision, sensitivity, specificity, F1-score, balanced accuracy, ROCAUC, PR-AUC, and confusion matrices were reported. Across classifiers, baseline balanced accuracy was 0.580; corrected means ranged from 0.696 to 0.720. Correction increased sensitivity but generally reduced precision and specificity. ROS achieved the highest mean F1-score (0.388), while SMOTE achieved the highest mean PR-AUC (0.378). A Friedman comparison indicated heterogeneity among techniques, followed by Holm-adjusted paired Wilcoxon tests. No approach dominated every classifier or metric. Leakage-safe correction and multi-metric assessment are essential; accuracy alone is unsuitable for selecting defect predictors.

Robust adaptive feature selection for imbalanced software defect prediction

Software defect prediction (SDP) often faces challenges related to heterogeneous software metrics, classifier dependency, and severe class imbalance, which may limit the robustness and generalization of feature selection strategies. This study proposes an adaptive feature selection approach to construct structured and discriminative feature subsets that remain effective across diverse datasets and learning models. The proposed method first measures the relationship between each software metric and the defect label using absolute determination power and then applies an adaptive retention rule to iteratively retain features with stronger discriminative contribution. The evaluation was conducted on multiple public defect datasets using several classical machine learning classifiers. Unlike approaches optimized for specific classifiers, the proposed strategy emphasizes cross-classifier robustness and imbalance-aware evaluation through defect recall and Matthews correlation coefficient. Experimental results show that the proposed method achieves a competitive average MCC of 0.257 and a defect recall of 0.446 compared with baseline approaches, although the statistical tests do not indicate significant superiority. Therefore, the proposed method should be interpreted as a comparable and stable alternative for feature selection under imbalanced SDP conditions. Stability and statistical analyses further indicate that the proposed method maintains comparable performance across dataset-classifier combinations. In addition, feature compactness analysis shows that the performance gains are attributable to efficient, interpretable feature subsets, highlighting the importance of robustness-oriented feature selection in SDP.

Impact of AI-Driven Techniques in Software Impact Analysis

Software defect prediction (SDP) becomes extremely important in enhancing the quality of software. Recent progresses in machine learning and ensemble learning have led to a great improvement on the prediction model of defects. In this research, a test defect prediction model combining state of (XGBoost, LightGBM, CatBoost) and balanced ensemble (EasyEnsemble, RUSBoost, Balanced Random Forest). The model is assessed on 5 standard AEEEM benchmark problems (EQ, JDT, LC, ML, PDE) with SMOTE oversampling and on the hold-out test strategy. The results of the experiment show RUSBoost- based model is more effective than the past models of defect prediction on EQ data with an AUC of 0.946, CatBoost model has an AUC of 0.850 on the JDT dataset, XGBoost method that uses on the LC sample has an AUC of 0.782, which is better than classical and other machine-learning-based methods published before, RUSBoost model is superior to the traditional and deep-learning-based models, which have the largest AUC of 0.757 to date on the ML dataset, XGBoost classifier obtains the high performance on the PDE dataset, with an AUC of 0.816, which is better than the classical and neural network baselines.

A Comparative Study of CNN and MLP for Software Defect Prediction Using PROMISE Datasets

Modern Software defects pose significant challenges, leading to critical system failures and substantial financial losses. As contemporary software systems become increasingly large and complex, identifying defects during the early stages grows more difficult. To address this, deep learning techniques, specifically multi-layer perceptron (MLP) and Convolutional Neural Network (CNN), are employed to predict software defects (SDP), integrated with code segment analysis during early development phases. This study applies both algorithms to detect software defects using 11 open-source datasets from the PROMISE repository. The evaluation emphasizes the prediction accuracy of the MLP and CNN models, alongside the F1 score-a crucial metric for assessing model performance on imbalanced datasets. Findings indicate that CNN outperforms MLP, achieving 86% prediction accuracy and an F1 score of 87.7%. In contrast, MLP attained 71% accuracy with an F1 score of 71.6%. These results demonstrate the superior predictive capability of CNN-based approaches in software defect prediction tasks.