Research map / Label-efficient and robust learning / Recognition and long-tail
Provably Better Explanations with Optimized Aggregation of Feature Attributions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes optimal convex aggregation of feature attributions across methods via a quadratic program over generalized L2 quality metrics (robustness and faithfulness), with provable improvement and generalization bounds.
Problem
Individual feature attribution methods are often unstable, hyperparameter-sensitive and unfaithful, and there is no principled way to combine different attributions to improve explanation quality.
Approach
Cast explanation quality metrics (Average-Sensitivity, Infidelity, and others) as generalized L2 metrics so that the quality of a convex combination of attributions becomes a quadratic form in the weights; solve a constrained QP for weights, with provable improvement over the weighted individual qualities and an O(1/sqrt(m)) generalization bound.
Key results
- AGG_robust reduces Average-Sensitivity on VGG16 to 0.424 vs 0.599 for best individual method VarGrad (lower is better).
- AGG_faith achieves the lowest Infidelity on DenseNet121 (2.390 vs 2.604 for SmoothGrad) and on ResNet18 (2.595 vs 2.916).
- AGG_opt improves ROAD MoRF_10 on ResNet18 to -3.30 versus -2.37 for VarGrad, and AGG_robust is best at -3.36.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Optimize convex aggregation weights for Average-Sensitivity | Gradient-based attribution methods · ImageNet-ILSVRC2012, VGG16; AlexNet; ResNet18; MobileNetV2; MLPMixer | SENS_AVG: VGG16: 0.582 -> 0.424; AlexNet: 0.476 -> 0.349; ResNet18: 0.518 -> 0.410; MobileNetV2: 0.568 -> 0.505; MLPMixer: 0.631 - | Helped |
| Optimize convex aggregation weights for Infidelity | Feature attribution methods · ImageNet-ILSVRC2012, DenseNet121; ResNet18; MobileNetV2; DeiT; SwinT | INFD: DenseNet121: 2.514 -> 2.390; ResNet18: 2.729 -> 2.595; MobileNetV2: 0.467 -> 0.443; DeiT: 0.339 -> 0.335; SwinT: 0.341 - | Helped |
| Optimize aggregation jointly for Infidelity and Average-Sensitivity instead of optimizing faithfulness alone | Feature attribution methods · ImageNet-ILSVRC2012, DenseNet121 | INFD: 2.390 -> 2.514 | Hurt |
| Optimize aggregation weights for robustness | Feature attribution methods · ResNet18 | RIS: 9.47 -> 3.27 | Helped |
| Optimize aggregation weights for robustness | Feature attribution methods · ResNet18 | RRS: 6.08 -> 2.09 | Helped |
| Optimize aggregation weights for robustness | Feature attribution methods · ResNet18 | ROS: 10.28 -> 3.55 | Helped |
| Optimize aggregation weights for faithfulness using blurred-image perturbations | Feature attribution methods · ResNet18, alternative zero and image-mean corruptions | INFD: 3.21 -> 2.82 (INFD_0); 3.10 -> 2.78 (INFD_xbar) | Helped |
| Optimize aggregation jointly for faithfulness and robustness | Feature attribution methods · ResNet18, zero corruption | FCOR_0: 0.44 -> 0.49 | Helped |
| Optimize aggregation weights for robustness | Feature attribution methods · ResNet18, ROAD MoRF_10, MoRF_20, and MoRF_30 | ROAD: -3.09 -> -3.36; -4.86 -> -5.48; -6.31 -> -7.17 | Helped |
| Optimize aggregation weights for faithfulness | Feature attribution methods · CIFAR10, BloodMNIST, DermaMNIST, and PathMNIST | INFD: CIFAR10: 3.538 -> 3.342; BloodMNIST: 11.80 -> 10.66; DermaMNIST: 0.322 -> 0.281; PathMNIST: 5.135 -> 4.850 | Helped |
Surprising findings
- The optimal weights vary strongly across samples and differ across architectures, so no single attribution method is consistently favored.
- Aggregated explanations for LIME variants automatically favor the sparser variants when objects are small.
Limitations
- Higher inference time than a single attribution method.
- Relies on the validity and diversity of the underlying attribution methods; a uniform weakness can compromise aggregation.
- Only convex weighting is considered; more sophisticated ensembling is left for future work.
Experimental setup
Datasets: ImageNet ILSVRC2012 CIFAR10 BloodMNIST DermaMNIST PathMNIST
Benchmarks: ImageNet ILSVRC2012 CIFAR10 BloodMNIST DermaMNIST PathMNIST
Metrics: Average-Sensitivity Max-Sensitivity Infidelity Faithfulness correlation RIS RRS ROS ROAD
Baselines: AGG_Mean AGG_Var SmoothGrad VarGrad Integrated Gradients GradCAM
Builds on: Average-Sensitivity Infidelity SmoothGrad error ambiguity decomposition
Closest papers by meaning
- Selective Explanations (NeurIPS 2024 · similarity 0.89)
- How to Probe: Simple Yet Effective Techniques for Improving Post-hoc Explanations (ICLR 2025 · similarity 0.88)
- Feature Attribution with Necessity and Sufficiency via Dual-stage Perturbation Test for Causal Explanation (ICML 2024 · similarity 0.87)
- GEFA: A General Feature Attribution Framework Using Proxy Gradient Estimation (ICML 2025 · similarity 0.87)
- Efficient and Accurate Explanation Estimation with Distribution Compression (ICLR 2025 · similarity 0.86)
- Regional Explanations: Bridging Local and Global Variable Importance (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.