Dynamic stacking ensemble for cross-language code smell detection

Area of Science:

• Software Engineering
• Machine Learning
• Computer Science

Background:

• Code smells are indicators of poor software design and implementation.
• Machine learning for code smell detection is an active research area, but models often lack stability and focus primarily on Java.
• Existing methods for code smell detection can be complex and may not generalize well across programming languages.

Purpose of the Study:

• To propose dynamic ensemble methods for code smell detection in both Java and Python.
• To investigate the effectiveness of greedy search and backward elimination strategies in building these dynamic ensembles.
• To compare the complexity and detection performance of dynamic ensembles against full stacking ensembles.

Main Methods:

• Development of dynamic ensemble models using greedy search and backward elimination strategies.
• Evaluation of detection performance on four Java and two Python code smells.
• Comparative analysis of model complexity and detection accuracy with full stacking ensembles.

Main Results:

• Greedy search and backward elimination yielded distinct sets of base models for dynamic ensembles.
• Dynamic ensembles demonstrated comparable detection performance to full stacking ensembles with no significant loss.
• Dynamic ensembles, particularly those using backward elimination, resulted in less complex models for most investigated code smells.

Conclusions:

• Dynamic stacking ensembles provide an effective and stable method for detecting Java and Python code smells.
• These dynamic ensembles offer a reduced complexity alternative to full stacking ensembles.
• The proposed strategies facilitate improved code smell detection across multiple programming languages.