Abstract
Deep learning models, particularly Transformer-based architectures, have achieved state-of-the-art performance in time-series forecasting. However, their "black-box" nature poses significant challenges for high-stakes decision-making, such as public health policy. This paper introduces a framework to benchmark the reliability of perturbation-based sensitivity analysis (SA) methods. By evaluating these methods across multiple deep learning architectures using real-world COVID-19 demographic data, the study assesses the consistency, robustness, and accuracy of feature importance explanations against empirical ground truth.
Introduction
As deep learning models become increasingly prevalent in time-series forecasting, the need for interpretability grows. Policymakers require transparency to understand which input features—such as specific age demographics—drive model predictions. While various "post-hoc" interpretation methods exist, their reliability remains under-explored in the context of time-series data.
This research addresses three critical gaps in explainable AI (XAI):
- Validation: Determining if different explanation methods yield consistent results.
- Robustness: Assessing whether explanations remain stable across different model architectures.
- Accuracy: Verifying if the identified "important features" align with real-world evidence.
Methodology
The study focuses on Perturbation-Based Sensitivity Analysis (SA). These methods are model-agnostic, meaning they do not require access to the internal gradients of a model. Instead, they involve perturbing or masking input data and observing the resulting changes in output.
The experimental framework incorporates several key components, ranging from the specific SA techniques used to the diverse set of deep learning models evaluated.
Sensitivity Analysis Techniques
The research evaluates three primary SA techniques:
- Feature Ablation & Occlusion: These methods systematically remove or mask specific features to quantify their impact on the model's predictions.
- Morris Method: A global sensitivity analysis that varies one parameter at a time across a grid of points. This approach is particularly effective at capturing non-linear effects and interactions between features. The author proposes an extended scaled Morris index to improve feature ranking in time-series contexts.
Model Architectures
To test for robustness, the framework utilizes a suite of modern time-series models, including the Temporal Fusion Transformer (TFT), TimesNet, Autoformer, DLinear, and PatchTST.
Evaluation Metrics
The primary metric for comparing feature importance rankings across different methods and models is the Spearman Rank Correlation Coefficient. This allows for a quantitative assessment of how well different explanations agree with each other and with the ground truth.
Dataset and Ground Truth
The framework is applied to a high-stakes, real-world scenario: U.S. County-Level COVID-19 Cases. The dataset spans two years and covers 3,142 counties, featuring eight distinct population age group features.
A unique contribution of this work is the use of objective Ground Truth. Unlike many XAI studies that rely on subjective plausibility, this research uses actual reported COVID-19 cases aggregated by age group. This allows for a direct validation of whether the model’s sensitivity rankings accurately reflect real-world infection patterns.
Key Insights and Results
The research highlights several critical findings regarding the state of XAI in time-series forecasting:
- Objective Validation: By moving beyond subjective "plausibility" and utilizing demographic infection data as a benchmark, the study provides a path toward objective accuracy in model explanations.
- Model-Agnostic Versatility: The choice of perturbation-based methods ensures that the interpretability framework is versatile and applicable to diverse architectures without requiring internal model access.
- Consistency and Robustness: Preliminary results indicate that the framework effectively benchmarks whether different SA methods yield comparable rankings and whether these rankings are sensitive to changes in the underlying deep learning architecture.
- Methodological Advancement: The proposed extension of the Morris Method (the scaled Morris index) demonstrates improved effectiveness in ranking feature importance for time-series data.
Conclusion and Future Work
This paper establishes a rigorous framework for validating the reliability of sensitivity analysis in time-series deep learning. For AI to be trusted in government and policy-making, transparency must be grounded in objective reality rather than architectural artifacts.
Future research will aim to:
- Expand Methodological Scope: Include back-propagation (gradient-based) and approximation methods (e.g., LIME, SHAP) for a broader comparison.
- Cross-Domain Validation: Test the framework in other domains like finance or energy where ground truth may be more elusive.
- Uncertainty Quantification: Investigate how data noise and perturbations contribute to overall model uncertainty.
Summary updated