Research map / Probabilistic and scientific modeling / Molecules and proteins
ReX: A Framework for Incorporating Temporal Information in Model-Agnostic Local Explanation Techniques
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 ReX, a framework that adds temporal predicates and length-varying perturbations to local model-agnostic explainers (Anchors, LIME, Kernel SHAP) so they capture position effects in variable-length sequence models.
Problem
Local model-agnostic explanation techniques ignore temporal information, so their explanations have low fidelity for models that process variable-length sequential inputs such as RNNs, transformers, and LLMs.
Approach
Keep the core algorithms of existing local explainers but extend their predicate set with 1-D (absolute position) and 2-D (relative position) temporal predicates, and extend the perturbation model with feature deletion and swapping so samples of different lengths are generated.
Key results
- ReX improves average Anchors coverage by 218.7% while keeping precision roughly the same.
- ReX-augmented LIME and Kernel SHAP improve average accuracy by 26.6% and 26.3% on sentiment and anomaly tasks, and AUROC by 45.8% and 38.0%; both surpass DecompX on BERT.
- In a 19-user study, Anchors* reached 87.3% precision_u and 68.8% coverage_u versus 48.3% and 43.9% for Anchors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ReX temporal predicates and variable-length sampling to Anchors | Anchors · Sentiment analysis and anomaly detection | coverage: average coverage by 218.7% | Helped |
| Add ReX temporal predicates and variable-length sampling to LIME | LIME · Sentiment analysis and anomaly detection | explanation accuracy: 26.6% | Helped |
| Add ReX temporal predicates and variable-length sampling to Kernel SHAP | Kernel SHAP · Sentiment analysis and anomaly detection | explanation accuracy: 26.6% and 26.3% | Helped |
| Add ReX temporal predicates and variable-length sampling to LIME | LIME · Text generation, Llama 2 | MSE: decreases the average MSE by 0.133 | Helped |
| Add ReX temporal predicates and variable-length sampling to Kernel SHAP | Kernel SHAP · Text generation, Llama 2 | MSE: decreases the average MSE by 0.041 | Helped |
| Add both 1-D and 2-D temporal predicates (ablation) | LIME · Sentiment analysis | explanation fidelity | Helped |
| Limit temporal-predicate window to w=5 instead of w=∞ (ablation) | ReX-augmented LIME · Sentiment analysis, LSTM; BERT; T5; GPT-2 | accuracy (%): 73.93 -> 75.09; 69.45 -> 68.20; 87.74 -> 86.99; 68.66 -> 69.03 | Mixed |
| Limit temporal-predicate window to w=5 instead of w=∞ (ablation) | ReX-augmented LIME · Sentiment analysis, LSTM; BERT; T5; GPT-2 | AUROC: 0.893 -> 0.887; 0.929 -> 0.927; 0.926 -> 0.924; 0.726 -> 0.759 | Mixed |
| Limit temporal-predicate window to w=5 instead of w=∞ (ablation) | ReX-augmented LIME · Sentiment analysis, LSTM; BERT; T5; GPT-2 | execution time (seconds): 24.52 -> 3.98; 23.42 -> 5.04; 27.13 -> 7.41; 28.39 -> 8.27 | Helped |
| Use ReX without 1-D temporal predicates (ablation) | LIME · Sentiment analysis | explanation fidelity | Hurt |
| Use ReX without 2-D temporal predicates (ablation) | LIME · Sentiment analysis | explanation fidelity | Hurt |
Surprising findings
- ReX-augmented Anchors often ran faster than vanilla Anchors, despite more predicates.
Limitations
- Requires realistic perturbation models, which are hard to define for domains like stock prices.
- Increases the number of predicates and regression cost for LIME and Kernel SHAP.
- User study is small (19 undergraduates) and limited to one model.
Experimental setup
Datasets: Stanford Sentiment Treebank ECG dataset Google Natural Questions
Benchmarks: Stanford Sentiment Treebank ECG anomaly detection Google Natural Questions
Metrics: precision coverage accuracy AUROC MSE execution time user study precision_u and coverage_u
Baselines: Anchors LIME Kernel SHAP DecompX C-LIME L-SHAP
Builds on: Anchors LIME Kernel SHAP MExGen
Closest papers by meaning
- Explainable Neural Networks with Guarantee: A Sparse Estimation Approach (AAAI 2025 · similarity 0.85)
- Optimal Information Retention for Time-Series Explanations (ICML 2025 · similarity 0.85)
- Faithful and Efficient Explanations for Neural Networks via Neural Tangent Kernel Surrogate Models (ICLR 2024 · similarity 0.85)
- A Neural Network Approach for Efficiently Answering Most Probable Explanation Queries in Probabilistic Models (NeurIPS 2024 · similarity 0.85)
- Explaining, Fast and Slow: Abstraction and Refinement of Provable Explanations (ICML 2025 · similarity 0.84)
- Data-centric Prediction Explanation via Kernelized Stein Discrepancy (ICLR 2025 · similarity 0.84)
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.