[Paper] The Directed Prediction Change - Efficient and Trustworthy Fidelity Assessment for Local Feature Attribution Methods

Source: arXiv - 2511.21363v1

Overview

The paper introduces Directed Prediction Change (DPC), a new fidelity metric for local feature‑attribution explanations. By refining the classic Prediction Change (PC) metric, DPC removes the heavy Monte‑Carlo sampling of Infidelity, delivering deterministic, ten‑times faster evaluations while still measuring how faithfully an explanation mirrors the model’s decision logic.

Key Contributions

• Directed Prediction Change (DPC): a deterministic fidelity metric that incorporates the sign of both perturbations and attributions, eliminating random sampling.
• Speedup: Achieves roughly a 10× reduction in computation time compared with the widely used Infidelity metric.
• Broad Empirical Validation: Tested on 4,744 explanations across two domains (dermatology image classification & financial tabular risk scoring), two black‑box models, seven attribution algorithms, and many hyper‑parameter settings.
• Holistic Evaluation Framework: Shows that combining DPC with the original PC metric provides a complete picture of both baseline‑oriented and local attribution quality.
• Reproducibility: The deterministic nature of DPC ensures that repeated runs yield identical results, a crucial property for regulatory and audit contexts.

Methodology

1. Guided Perturbation Experiment – The authors start from the existing PC setup, where a small perturbation is added to an input and the resulting change in the model’s prediction is measured.
2. Incorporating Direction – Instead of treating perturbations as unsigned magnitudes, DPC multiplies each perturbation by the sign of the corresponding attribution value. This aligns the perturbation with the explanation’s claimed influence (positive or negative).
3. Deterministic Computation – Because the perturbation direction is now fixed by the attribution, there is no need to draw random samples to estimate an expectation. The metric can be computed with a single forward pass per perturbed feature.
4. Comparison to Infidelity – The authors analytically demonstrate that DPC approximates the same underlying fidelity property that Infidelity captures, but without the stochastic estimator.
5. Experimental Protocol – They evaluate DPC on:
   • Datasets: ISIC skin‑lesion images (CNN) and a credit‑risk tabular dataset (gradient‑boosted trees).
   • Attribution Methods: Grad‑CAM, Integrated Gradients, SHAP, LIME, etc.
   • Hyper‑parameters: Varying perturbation magnitudes, number of perturbed features, and attribution smoothing settings.

Results & Findings

• Speed: DPC runs in ~0.1 s per explanation versus ~1 s for Infidelity (≈10× faster).
• Correlation: DPC scores correlate strongly (Pearson ≈ 0.92) with Infidelity, confirming they measure the same fidelity aspect.
• Stability: Unlike Infidelity, DPC yields identical scores across repeated runs, removing variance caused by Monte‑Carlo sampling.
• Discriminative Power: DPC successfully differentiates high‑fidelity explanations (e.g., Integrated Gradients) from noisy ones (e.g., vanilla gradient) across both image and tabular domains.
• Complementarity: When paired with PC, DPC highlights cases where an explanation captures directional influence (DPC) but not magnitude (PC), offering richer diagnostic insight.

Practical Implications

• Faster Model Audits – Teams can now assess thousands of explanations in minutes, enabling continuous monitoring of model interpretability pipelines.
• Regulatory Compliance – Deterministic, reproducible fidelity scores simplify documentation for medical device regulators or financial auditors who demand audit trails.
• Tooling Integration – DPC can be added as a lightweight plug‑in to existing XAI libraries (e.g., Captum, Alibi) without requiring extra sampling infrastructure.
• Guided Model Debugging – Developers can quickly spot attribution methods that misrepresent model behavior, leading to more reliable debugging and feature engineering.
• Resource‑Constrained Environments – Because DPC needs only a single forward pass per perturbed feature, it is suitable for on‑device or edge‑AI scenarios where compute budgets are tight.

Limitations & Future Work

• Perturbation Scope – DPC still relies on a predefined perturbation magnitude; choosing an inappropriate scale could mask fidelity issues.
• Local vs. Global – The metric evaluates fidelity locally around a single instance; extending it to capture global explanation quality remains open.
• Model Types – Experiments focus on CNNs and gradient‑boosted trees; assessing DPC on transformer‑based or reinforcement‑learning models would broaden its applicability.
• Attribution Diversity – While seven methods were tested, newer hybrid or counterfactual explanations were not included; future studies could explore DPC’s behavior on those.

Bottom line: Directed Prediction Change offers a practical, fast, and reproducible way for developers and data scientists to verify that their model explanations truly reflect the underlying decision logic—an essential step toward trustworthy AI in high‑stakes applications.

Authors

• Kevin Iselborn
• David Dembinsky
• Adriano Lucieri
• Andreas Dengel

Paper Information

• arXiv ID: 2511.21363v1
• Categories: cs.LG, cs.AI
• Published: November 26, 2025
• PDF: Download PDF