Uncertainty-aware sensitivity analysis using Rényi divergences

Topi Paananen, Michael Riis Andersen, Aki VehtariUAI 2021

Abstract

Standard sensitivity analysis typically measures how a model's mean prediction changes with respect to its inputs. However, this approach ignores the model's predictive uncertainty, which can lead to misleading results in regions with sparse data or model bias. We present R-sens, a novel uncertainty-aware sensitivity measure based on the curvature of Rényi divergences. By accounting for the full predictive distribution, R-sens provides more reliable feature importance rankings and interaction detection. Our experiments on synthetic and real-world datasets demonstrate that R-sens is more stable than traditional methods and identifies interactions that significantly improve out-of-sample predictive performance.

Introduction

In machine learning, sensitivity analysis is crucial for interpreting complex models and identifying key features. The standard approach involves calculating the derivative of the model's output with respect to its inputs. While intuitive, this method treats the model as a deterministic function, ignoring the epistemic uncertainty inherent in probabilistic models.

A high derivative in a region where the model has little data may simply reflect noise or extrapolation artifacts rather than true feature importance. Conversely, an input might significantly affect the model's confidence (variance) without changing the mean prediction—an effect that standard sensitivity analysis completely misses. To address these limitations, we propose a framework that incorporates the entire predictive distribution into the sensitivity measure.

Comparison of standard sensitivity (black) and uncertainty-aware R-sens (red).

Methodology: R-sens and R-sens₂

The core idea behind R-sens is to measure sensitivity by differentiating the Rényi divergence between the predictive distribution at a point $x$ and a perturbed point $x + \Delta x$.

The Curvature of Divergence

By taking the limit as the perturbation $\Delta x$ approaches zero, we focus on the curvature (second derivative) of the divergence. For many common distributions, this curvature is related to the Fisher Information Matrix (FIM) of the predictive distribution. This connection allows R-sens to weight the derivatives of model parameters by the information they carry. For a Gaussian predictive distribution, R-sens considers both the derivative of the mean and the derivative of the variance.

Detecting Interactions

We extend this framework to identify pairwise interactions through R-sens₂. By examining higher-order derivatives of the divergence, R-sens₂ can detect variables that interact not just in their effect on the mean, but also in their effect on the model's uncertainty. This provides a more holistic view of feature interactions than traditional methods like the H-statistic or SHAP.

Experimental Evaluation

We validated R-sens using synthetic benchmarks and two real-world datasets: the Concrete Slump dataset (predicting compressive strength) and the Bike Sharing dataset (predicting rental demand).

In simulations with biased models, R-sens maintained accurate feature rankings even when the model was intentionally mis-specified in low-data regions. Standard derivative-based methods often failed in these scenarios, being misled by high-variance gradients in the extrapolation regions. Furthermore, R-sens demonstrated significantly higher stability across bootstrap samples compared to popular methods like SHAP or Permutation Feature Importance.

Real-World Performance

On the Concrete and Bike Sharing datasets, we compared R-sens₂ against popular interaction detection methods. We evaluated these methods by identifying the top $k$ interactions and adding them to a predictive model, measuring the resulting improvement in out-of-sample log predictive density.

Predictive performance (MLPD) on Concrete and Bike sharing datasets for different interaction detection methods.

As shown in the figure above, models incorporating interactions identified by R-sens₂ consistently achieved higher Mean Log Predictive Density (MLPD) compared to those using interactions from other methods. This suggests that R-sens₂ captures more meaningful and predictive relationships within the data.

Conclusion

R-sens offers a robust, uncertainty-aware alternative to traditional sensitivity analysis. By leveraging the geometry of predictive distributions through Rényi divergences, it provides a principled way to filter out unreliable signals in sparse data regions and capture sensitivity in the model's uncertainty. While currently focused on models with analytical predictive distributions, such as Gaussian Processes and Bayesian Neural Networks, future work will explore extensions to non-parametric settings and different divergence families.

Summary updated

BibTeX
@inproceedings{paananen2021uncertainty,
  author = {Topi Paananen and Michael Riis Andersen and Aki Vehtari},
  title = {{Uncertainty-aware sensitivity analysis using Rényi divergences}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2021},
}
Figure