Scholarlink

<p>Software reliability prediction and defect detection are essential for improving software quality and reducing development costs. Early and accurate defect prediction supports timely interventions and leads to better project outcomes. This study presents a hybrid approach that combines ensemble learning with the Synthetic Minority Oversampling Technique (SMOTE) and a Modified Jaya Optimization Algorithm (MJOA) for feature selection. The MJOA uses adaptive control parameters that improve convergence and prevent the issue of local optima, making the feature selection process more effective. A brief comparison shows that Greedy Feature Selection (GFS) is faster but often becomes trapped in local optima, whereas the Modified Jaya Optimization method provides stronger global search ability and selects more relevant features for accurate defect prediction. The proposed framework uses eight classifiers, namely Extreme Learning Machine (ELM), Random Forest (RF), Weighted Support Vector Machine (W SVM), Gradient Boosting (GB), Logistic Regression (LR), Stochastic Gradient Descent (SGD), Bernoulli Naive Bayes (BNB), and Multinomial Naive Bayes (MNB), combined through average probability voting. This ensemble method leverages the strengths of individual classifiers, reduces variance and over fitting, and improves generalization performance. Experiments were carried out on six PROMISE datasets, specifically PC2, PC4, MC1, KC3, Camel 1.6, and Ant 1.7, using tenfold cross validation. The Jaya based ensemble performed better than baseline models as well as those using GFS and achieved up to 0.97 percent accuracy, an F measure of 0.95, and a precision of 0.93 on benchmark datasets.These results demonstrate the robustness and practical applicability of the proposed approach, and future work may investigate the integration of deep learning models and alternative optimization techniques to further improve scalability and performance.The study is limited by its use of PROMISE datasets, which may reduce generalization to large industrial settings, and by the higher computational cost of optimization based feature selection. Future work may explore deep learning methods and alternative optimizers to improve scalability and performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Software reliability prediction using ensemble learning and SMOTE on features selected by jaya optimization

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Suneel Kumar Rath,
  • Madhusmita Sahu,
  • Sukant Kishoro Bisoy,
  • Prabodh Kumar Sahoo,
  • Shom Prasad Das

摘要

Software reliability prediction and defect detection are essential for improving software quality and reducing development costs. Early and accurate defect prediction supports timely interventions and leads to better project outcomes. This study presents a hybrid approach that combines ensemble learning with the Synthetic Minority Oversampling Technique (SMOTE) and a Modified Jaya Optimization Algorithm (MJOA) for feature selection. The MJOA uses adaptive control parameters that improve convergence and prevent the issue of local optima, making the feature selection process more effective. A brief comparison shows that Greedy Feature Selection (GFS) is faster but often becomes trapped in local optima, whereas the Modified Jaya Optimization method provides stronger global search ability and selects more relevant features for accurate defect prediction. The proposed framework uses eight classifiers, namely Extreme Learning Machine (ELM), Random Forest (RF), Weighted Support Vector Machine (W SVM), Gradient Boosting (GB), Logistic Regression (LR), Stochastic Gradient Descent (SGD), Bernoulli Naive Bayes (BNB), and Multinomial Naive Bayes (MNB), combined through average probability voting. This ensemble method leverages the strengths of individual classifiers, reduces variance and over fitting, and improves generalization performance. Experiments were carried out on six PROMISE datasets, specifically PC2, PC4, MC1, KC3, Camel 1.6, and Ant 1.7, using tenfold cross validation. The Jaya based ensemble performed better than baseline models as well as those using GFS and achieved up to 0.97 percent accuracy, an F measure of 0.95, and a precision of 0.93 on benchmark datasets.These results demonstrate the robustness and practical applicability of the proposed approach, and future work may investigate the integration of deep learning models and alternative optimization techniques to further improve scalability and performance.The study is limited by its use of PROMISE datasets, which may reduce generalization to large industrial settings, and by the higher computational cost of optimization based feature selection. Future work may explore deep learning methods and alternative optimizers to improve scalability and performance.