Investigating the Accuracy of Metric-Based versus Machine Learning Approaches in Detecting Design Patterns

Abstract

Design pattern detection approaches have evolved, with machine-learning methods gaining prominence. However, implementing machine-learning models can be challenging due to extensive training requirements and the need for large labeled design pattern datasets. This study tests a simpler alternative that overcomes these specific machine learning limitations, and compares design pattern detection accuracy of machine-learning approaches and a metric approach, using both Java and C++. Without relying on AI, the metric approach achieves comparable or better fscore than existing machine learning methods by means of extracting metrics from programs using scripts. The findings demonstrate the potential of metric approaches as practical alternatives, simplifying design pattern analysis in software development. Future research should explore the application of metric approaches in industry contexts.