Scholarlink

In recent years, machine learning has become part of modern digital products. However, ensuring the quality of these systems involves challenges that go well beyond conventional software code, touching aspects such as data integrity, pipeline reliability, and the social dynamics of development teams. This review explores how the research community has approached these issues through the notions of smells and antipatterns in ML-based software. Following the PRISMA-ScR methodology, we examined studies published between 2015 and 2025 in five major digital libraries. From an initial pool of 220 documents, only 9 met the inclusion criteria after detailed screening and forward snowballing. Their insights allowed the identification of 118 distinct smells and antipatterns grouped into four categories: data smells (51), community smells (12), ML code smells (41), and project-level antipatterns (14). The evidence shows that ML systems face quality risks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Software Quality in Machine Learning Systems: A Scoping Review of Smells and Antipatterns

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Marcela Mosquera,
  • Rodolfo Bojorque,
  • Remigio Hurtado

摘要

In recent years, machine learning has become part of modern digital products. However, ensuring the quality of these systems involves challenges that go well beyond conventional software code, touching aspects such as data integrity, pipeline reliability, and the social dynamics of development teams. This review explores how the research community has approached these issues through the notions of smells and antipatterns in ML-based software. Following the PRISMA-ScR methodology, we examined studies published between 2015 and 2025 in five major digital libraries. From an initial pool of 220 documents, only 9 met the inclusion criteria after detailed screening and forward snowballing. Their insights allowed the identification of 118 distinct smells and antipatterns grouped into four categories: data smells (51), community smells (12), ML code smells (41), and project-level antipatterns (14). The evidence shows that ML systems face quality risks.