Delta-Audit: Explaining What Changes When Models Change

2025/08/27 by Hemmat, Arshia, Fatemi, Afsaneh · 1 citation
#FOS: Computer and information sciences #Machine Learning (cs.LG)

paper · doi:10.48550/arxiv.2508.19589

Abstract

Model updates (new hyperparameters, kernels, depths, solvers, or data) change performance, but the reason often remains opaque. We introduce Delta-Attribution (Δ-Attribution), a model-agnostic framework that explains what changed between versions A and B by differencing per-feature attributions: Δϕ(x)=ϕ_B(x)-ϕ_A(x). We evaluate Δϕ with a Δ-Attribution Quality Suite covering magnitude/sparsity (L1, Top-k, entropy), agreement/shift (rank-overlap@10, Jensen--Shannon divergence), behavioural alignment (Delta Conservation Error, DCE; Behaviour--Attribution Coupling, BAC; COΔF), and robustness (noise, baseline sensitivity, grouped occlusion). Instantiated via fast occlusion/clamping in standardized space with a class-anchored margin and baseline averaging, we audit 45 settings: five classical families (Logistic Regression, SVC, Random Forests, Gradient Boosting, kNN), three datasets (Breast Cancer, Wine, Digits), and three A/B pairs per family. Findings. Inductive-bias changes yield large, behaviour-aligned deltas (e.g., SVC poly →rbf on Breast Cancer: BAC≈0.998, DCE≈6.6; Random Forest feature-rule swap on Digits: BAC≈0.997, DCE≈7.5), while ``cosmetic'' tweaks (SVC gamma=scale vs. auto, kNN search) show rank-overlap@10=1.0 and DCE≈0. The largest redistribution appears for deeper GB on Breast Cancer (JSD≈0.357). Δ-Attribution offers a lightweight update audit that complements accuracy by distinguishing benign changes from behaviourally meaningful or risky reliance shifts.

Citations

• Reliable Evaluation of Attribution Maps in CNNs: A Perturbation-Based Approach
• Rebuilding ROME : Resolving Model Collapse during Sequential Model Editing
• Explanation Shift: How Did the Distribution Shift Impact the Model?
• Explanation Shift: Detecting distribution shifts on tabular data via the explanation space
• "Why did the Model Fail?": Attributing Model Performance Changes to Distribution Shifts
• Mass-Editing Memory in a Transformer
• Locating and Editing Factual Associations in GPT
• Fast Model Editing at Scale
• An Information-theoretic Approach to Distribution Shifts
• Smoothed Geometry for Robust Attribution
• From local explanations to global understanding with explainable AI for trees
• Fooling LIME and SHAP: Adversarial Attacks on Post hoc Explanation\n Methods
• Sanity Checks for Saliency Maps
• A Benchmark for Interpretability Methods in Deep Neural Networks
• SmoothGrad: removing noise by adding noise
• A Unified Approach to Interpreting Model Predictions
• Axiomatic Attribution for Deep Networks
• Grad-CAM: Visual Explanations from Deep Networks via Gradient-Based Localization
• Visualizing and Understanding Convolutional Networks
• Divergence measures based on the Shannon entropy

Cited by

• Assessing Policy Updates: Toward Trust-Preserving Intelligent User Interfaces

Related

• The Directed Prediction Change - Efficient and Trustworthy Fidelity Assessment for Local Feature Attribution Methods
• Rethinking Distribution Shifts: Empirical Analysis and Inductive Modeling for Tabular Data
• Explainable AI Through a Democratic Lens: DhondtXAI for D'Hondt-Projected Feature Attribution
• Explanation Shift: Detecting distribution shifts on tabular data via the explanation space
• Data AUDIT: Identifying Attribute Utility- and Detectability-Induced Bias in Task Models