Beyond Feature Importance: Investigating Predictive Sensitivity in Machine Learning Models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Feature importance is widely used in machine learning to assess the contribution of individual predictors to model performance and support the interpretation of model behaviour. However, it remains unclear whether features identified as highly important are also those to which a model is most sensitive when their values are disturbed. This study empirically investigates the relationship between feature importance and predictive sensitivity under controlled feature perturbations. Experiments are conducted on five classification datasets, namely Sonar, Ionosphere, Breast Cancer Wisconsin, Parkinson’s Disease, and Spambase, using Logistic Regression, Random Forest, and Gradient Boosting to represent different modelling characteristics. Permutation importance is used to estimate model-specific feature importance. Individual features in the test data are then progressively perturbed using Gaussian noise at multiple intensity levels, while the trained models remain unchanged. Predictive sensitivity is measured through the resulting degradation in F1 score, and Spearman rank correlation is used to examine the correspondence between feature importance and feature-wise sensitivity. The results show that higher feature importance is associated with greater perturbation sensitivity in several dataset and model combinations. However, the relationship is not consistent across all experimental conditions or classifiers. Both the magnitude and statistical significance of the observed associations vary across datasets, models, and perturbation intensities, with stronger associations observed at higher perturbation levels in several cases. These findings suggest that feature importance and perturbation sensitivity capture related but distinct aspects of model behaviour. Therefore, feature importance alone may not fully reflect how strongly model performance responds to disturbances in individual features. Controlled feature perturbation analysis can thus provide a complementary perspective for interpreting feature relevance and evaluating the robustness of machine learning models.