Abstract
Variable importance is a cornerstone of machine learning interpretability, yet standard permutation importance (PI) often yields inflated false positive rates when features are correlated. We present Conditional Permutation Importance (CPI), a model-agnostic framework that leverages conditional sampling to provide statistically valid p-values and accurate feature rankings. By decoupling a feature into its predictable and residual components and permuting only the latter, CPI ensures that synthetic data points remain within the original distribution. Extensive evaluations on synthetic benchmarks and high-dimensional datasets like the UK Biobank demonstrate that CPI-DNN maintains robust Type-I error control and superior ranking performance compared to existing methods, all while remaining computationally efficient.
Introduction
In fields such as biomedicine, understanding why a model makes a prediction is as critical as the prediction itself. The traditional "gold standard" for this is Permutation Importance (PI), which measures the drop in model performance after shuffling a feature's values. However, PI suffers from a significant flaw: when features are correlated (e.g., height and weight), independent shuffling creates "impossible" data points. Models trained on realistic data perform poorly on these outliers, leading to inflated importance scores for redundant variables. This "correlation trap" results in high Type-I error rates, causing researchers to pursue false-positive biomarkers.
Methodology
To address these limitations, we propose Conditional Permutation Importance (CPI). Instead of global shuffling, CPI shuffles a variable $x^j$ conditionally on all other variables.
- Conditional Sampling: The variable of interest is split into a part predictable by other features and a residual part. Only the residuals are permuted, ensuring the resulting data respects the underlying correlation structure.
- Learner Agnostic Framework: CPI can be paired with any base learner. The authors emphasize CPI-DNN (Deep Neural Networks) for its high predictive capacity.
- Computational Efficiency: Unlike "Leave-One-Covariate-Out" (LOCO) methods that require full model retraining for every feature, CPI uses a lean Random Forest to estimate conditional distributions, significantly reducing overhead.
- Statistical Guarantees: The framework derives a Wald statistic for importance scores that follows a standard normal distribution in large samples, enabling the calculation of reliable p-values for feature selection.
Experimental Evaluation
The framework was evaluated across diverse scenarios:
- Synthetic Benchmarks: 100 simulations with correlation levels ($\rho$) ranging from 0 to 0.8 across five data-generating scenarios (e.g., Plain Linear, Interactions Only).
- UK Biobank (UKBB): A dataset of 8,357 individuals used to predict cognitive and mood characteristics from brain imaging and socio-demographic data.
- Cam-CAN (MEG): A dataset of 536 individuals for age prediction using over 4,000 magnetoencephalography features.
Key Results and Discussion
Robustness to Correlation
In synthetic tests, standard methods like Permfit-DNN saw false positive rates explode as correlation increased. In contrast, CPI-DNN maintained the targeted 5% Type-I error rate even at high correlation levels ($\rho=0.8$).
Comparative Performance
CPI-DNN consistently achieved the highest Area Under the Curve (AUC) scores for feature ranking across all scenarios, outperforming competitors such as d0CRT, Lazy VI, and LOCO. While LOCO methods provide similar statistical rigor, CPI-DNN is orders of magnitude faster because it avoids the need for repeated model refitting.
Scientific Discovery in Real-World Data
On the UK Biobank dataset, CPI demonstrated its practical utility by rejecting 16 brain-related variables that standard permutation methods had incorrectly labeled as significant. This provides a more parsimonious and scientifically rigorous map of the features actually driving cognitive health predictions, filtering out spurious correlations inherent in high-dimensional biological data.
Conclusion
CPI provides a scalable and statistically sound solution to the problem of variable importance in the presence of correlation. While it requires larger datasets for stable results when using DNNs, its ability to control Type-I errors makes it an essential tool for reliable feature discovery. Future work will explore extending this framework to perform inference on groups of variables, such as entire gene pathways, to further enhance interpretability in complex biological systems.
Summary updated