An Evidence-Based Study on the Relationship of Software Engineering Practices on Code Smells in Python ML Projects

The rapid adoption of Machine Learning (ML) technologies has introduced new challenges for code quality. Code smells, i.e., suboptimal design and implementation choices applied when developing source code, represent a particularly prevalent problem. While software engineering (SE) practices are often recommended to improve maintainability, their actual impact on code smells in ML projects remains unclear. In this paper, we present an evidence-based empirical study of 566 real-world Python ML projects from the NICHE dataset, labeled according to adherence to eight established SE practices. Using static analysis and statistical testing, we assess the relationship between these practices and the presence of ten Python-specific code smells. Our results show that projects adopting SE practices exhibit significantly fewer code smells. In particular, Continuous Integration is negatively correlated with the Complex Container Comprehension smell. These findings highlight the importance of engineering discipline in managing code quality in ML development.