Research Artefact:

Do Machine Learning Models Perform Consistently in Code Smell Detection? Analyzing Code Smell Trends in Python Repositorie

This repository implements a machine learning–based approach for detecting code smells in Python software projects. The goal is to improve traditional rule-based and threshold-based detection methods by leveraging software metrics and supervised learning models.

Overview

Code quality often degrades due to time pressure and changing requirements, leading to technical debt. One common indicator of technical debt is code smell, which reflects poor design or inefficient implementation.

This project focuses on automatically detecting code smells using machine learning and analyzing their distribution and evolution across large-scale Python repositories.

Code Smells Studied

The following four types of code smells are analyzed:

• Long Method
• Large Class
• Data Class
• Complex Method

Dataset

The dataset is constructed from four popular open-source Python repositories on Github:

• Django
• Matplotlib
• NumPy
• SciPy

Extracted Features

Each source code entity is represented using software metrics, including:

• Raw Metrics
• Cyclomatic Complexity
• Halstead Metrics
• Object Oriented Matrics

Machine Learning Models

The following models are implemented and evaluated:

• Logistic Regression
• Random Forest
• K-Nearest Neighbor (KNN)
• Multi-Layer Perceptron (MLP)
• TabNet

Evaluation Method

Model performance is evaluated using:

• Stratified 5-Fold Cross-Validation
• Accuracy
• Matthews Correlation Coefficient (MCC)

To assess statistical significance, the following tests are applied:

• Friedman Test
• Nemenyi Post-hoc Test

Analysis

In addition to model comparison, this project analyzes:

• The distribution of code smells across repositories
• The evolutionary trends of code smells over time

This analysis provides insights into how code quality changes during software evolution.

Research Questions

This study is guided by the following research questions:

• RQ1: How does hyperparameter tuning affect the performance of different machine learning models in detecting code smells?

• RQ2: How computationally efficient are different machine learning models in terms of training time and prediction time?

• RQ3: Which machine learning model is most suitable for detecting each type of code smell based on predictive performance and computational efficiency?

• RQ4: How are code smells distributed, and how do their trends evolve across different Python repositories when using the best-performing machine learning model?

Keywords

code smell, machine learning, software engineering, Python, GitHub