Determine the individual performance effects of ML-specific violations

Quantify the individual contribution of each SonarQube ML-specific rule violation to machine-learning performance, beyond the aggregate association measured across all ML violations.

Background

The study analyzes 34 SonarQube rules targeting data-science and machine-learning practices and finds that the aggregate number of ML-specific violations has a small negative association with Kaggle competition performance. However, violations differ substantially in prevalence, severity, and the mechanisms through which they may affect an ML pipeline.

The authors explicitly distinguish prevalence from impact: a frequently violated rule is not necessarily the rule most strongly associated with lower performance. The unresolved task is therefore to isolate the contribution of each individual ML-specific rule, which would clarify which practices matter most for model outcomes and sharpen the interpretation of the aggregate effect.

References

Finally, prevalence does not imply impact: the most frequently violated rules are not necessarily those most strongly associated with lower performance. Our study measures the aggregate association between ML violations and performance (RQ2). Quantifying the individual contribution of each rule is left as future work (Section~\ref{sec:discussion:researchers}).

— On the Relation between Code Quality and Machine Learning Performance: A Large-scale Empirical Study  (2609.10610 - Mignard et al., 8 Sep 2026) in Section 6.5, subsection “Our most violated ML rules”; referenced Section 7.3, subsection “For researchers”

The diagnostics characterize that regime, while the relation between UVI values and task quality is an empirical question.

— Neural Dimensional Analysis: Foundations of Dimensionless and Scale-Consistent Computation  (2610.17969 - Kim et al., 13 Oct 2026) in Section 'Discussion', paragraph 'The status of a unit declaration'