Research map / Time series / Foundation models for time series
InteDisUX: Intepretation-Guided Discriminative User-Centric Explanation for Time Series
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 InteDisUX, a segment-based time series explanation method that starts from equal segments scored by segment-level integrated gradients and greedily merges neighbors using discrimination and faithfulness gains.
Problem
Explaining deep time series classifiers with segment-level explanations that are faithful and robust, avoiding the fixed pre-cut segmentation and LIME instability, poor local fidelity, and imbalance issues of existing LIME-based methods.
Approach
Replaces LIME scoring with segment-level integrated gradients (SIG) and builds the segmentation bottom-up by greedily merging adjacent segments, guided by a discrimination gain (Hausdorff distance between consecutive segments) and a model-based faithfulness gain, so the segmentation itself uses model information.
Key results
- InteDisUX's faithfulness exceeds LIMESegment on all 12 UCR datasets, reaching up to 0.487 gain as stated in the text (e.g., Chinatown 0.776 vs 0.333, Coffee 0.578 vs 0.348).
- InteDisUX shows better robustness than LIMESegment on 8 of 12 datasets (e.g., Hand 0.757 vs 0.597, Yoga 0.456 vs 0.317).
- Ablations show SIG improves faithfulness over LIME in 11/12 datasets, and the faithfulness gain improves faithfulness over DisSIGSegment on all 12 datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use InteDisUX instead of LIMESegment | Time-series explanations · 12 UCR datasets | Faithfulness: outperforms ... across 12 datasets up to a score of 0.487 | Helped |
| Use InteDisUX instead of LIMESegment | Time-series explanations · 8/12 UCR datasets | Robustness: 8 out of 12 datasets | Helped |
| Replace LIMESegment's LIME scoring with SIG, keeping its segmentation strategy (ablation) | SIGSegment · 11/12 UCR datasets | Faithfulness: 11/12 datasets | Helped |
| Replace LIMESegment's LIME scoring with SIG, keeping its segmentation strategy (ablation) | SIGSegment · 5/12 UCR datasets | Robustness: 5/12 dataset | Mixed |
| Add discrimination-gain-guided merging to SIGSegment (ablation) | DisSIGSegment · 12 UCR datasets | Faithfulness: 8/12 datasets; 6/12 datasets | Mixed |
| Add discrimination-gain-guided merging to SIGSegment (ablation) | DisSIGSegment · 12 UCR datasets | Robustness: 8/12 datasets; 11/12 dataset | Helped |
| Add faithfulness-gain constraint to candidate selection (ablation) | InteDisUX · 12 UCR datasets | Faithfulness: all 12 datasets | Helped |
| Add faithfulness-gain constraint to candidate selection (ablation) | InteDisUX · 12 UCR datasets | Robustness | No clear effect |
| Increase the initial segment count Ns (ablation) | InteDisUX · Yoga | Faithfulness | Helped |
| Increase the initial segment count Ns (ablation) | InteDisUX · Yoga | Robustness | Hurt |
Surprising findings
- LIMESegment's robustness is higher than InteDisUX on several datasets (e.g., Strawberry 0.863 vs 0.514, GPAge/GPGender/Dodger near 0.98-0.99 for LIMESegment or InteDisUX), so robustness is not uniformly improved.
- SIG-based segment scoring improves faithfulness but only improves robustness over LIMESegment on 5/12 datasets, suggesting the gain comes largely from the new segmentation.
Limitations
- Robustness is not consistently better than LIMESegment (only 8 of 12 datasets).
- Faithfulness scores remain modest on several datasets (e.g., GPAge 0.249, Freezer 0.252), and the approach depends on a choice of initial segment count Ns whose effect varies.
Experimental setup
Datasets: Coffee Strawberry GunPointOldVersusYoung HandOutlines Yoga ECG200 GunPointMaleVersusFemale Dodger Chinatown FreezerSmallTrain HouseTwenty WormsTwoClass
Benchmarks: Coffee Strawberry GunPointOldVersusYoung HandOutlines Yoga ECG200 GunPointMaleVersusFemale Dodger Chinatown FreezerSmallTrain HouseTwenty WormsTwoClass
Metrics: faithfulness robustness
Baselines: LIMESegment Neves et al. 2021 LIME-based method Guillemé et al. 2019 LIME-based method
Builds on: Integrated Gradients LIMESegment LIME
Closest papers by meaning
- TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation (ICML 2025 · similarity 0.86)
- Optimal Information Retention for Time-Series Explanations (ICML 2025 · similarity 0.85)
- ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models (NeurIPS 2025 · similarity 0.85)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.84)
- SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (NeurIPS 2025 · similarity 0.84)
- TimeXL: Explainable Multi-modal Time Series Prediction with LLM-in-the-Loop (NeurIPS 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.