Alleviating class imbalance in Feature Envy prediction: An oversampling technique based on code entity attributes

Abstract

Context:

Feature Envy is a common code smell that occurs when a method heavily relies on data or functionality from other classes. Detecting Feature Envy is essential for improving software modularity and reducing technical debt. However, real-world datasets often exhibit severe class imbalance, with far fewer Feature Envy instances than non-smelly ones, posing challenges for prediction models. Traditional oversampling techniques attempt to address this issue by relying solely on numerical vectors but often fail to capture the complex relationships between code entities, potentially deviating from the nature of Feature Envy.

Objective:

This study introduces STANDER, a novel oversampling technique based on code entity similarity, designed to handle class imbalance in Feature Envy prediction by generating synthetic samples that better reflect the characteristics of Feature Envy.

Method:

STANDER creates synthetic samples by leveraging multidimensional code entity similarity, incorporating attributes such as dependency relationships, historical changes and code text. It was evaluated on five datasets using five classifiers: Naive Bayes, Logistic Regression, Support Vector Machine, Random Forest, and Decision Tree. Its performance was compared to baseline over-sampling techniques based on precision, recall, F1-score, and Matthews Correlation Coefficient.

Results:

STANDER enhances dataset diversity while maintaining clear boundaries between minority and majority classes, as reflected by higher Nearest Neighbor Diversity and Silhouette Score values. Models balanced with STANDER exhibited significant improvements in predictive performance, particularly in recall, F1-score, and Matthews Correlation Coefficient. Compared to the other oversampling techniques, STANDER demonstrated advantages in handling imbalanced datasets, especially in the Logistic Regression and Decision Tree classifiers. Statistical results confirm significant performance improvements across most models, highlighting its effectiveness and applicability.

Conclusion:

STANDER is an effective solution to alleviate class imbalance problem in Feature Envy detection by generating more representative synthetic samples that improve prediction performance.