Research map / Time series / Causal and generative time series
TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation
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 cumulative metrics (CPD, CPP) that expose the cancel-out flaw in simultaneous-masking evaluations, and TIMING, a segment-masked Integrated Gradients variant for time series attribution that outperforms baselines on synthetic and real datasets.
Problem
Time series XAI methods mostly rank points by unsigned attribution magnitude, and existing evaluations that mask top-K points simultaneously cancel out opposing-sign contributions, so directional methods like IG are underrated; IG on time series also produces OOD interpolation paths and ignores temp
Approach
Introduce cumulative metrics (CPD, CPP) that measure prediction changes step by step rather than with simultaneous masking, and enhance IG by replacing the zero baseline with random segment-based retention masks so intermediate path points stay closer to the data distribution while keeping IG-like properties.
Key results
- On MIMIC-III with zero substitution, TIMING reaches CPD (K=50) 0.366 vs 0.342 for IG and 0.204 for Extrmask (cumulative masking).
- TIMING improves zero-substitution CPD over the best prior method by relative gains of 8.7% on MIMIC-III, 5.1% on PAM, 109.8% on Boiler, 11.1% on Epilepsy, 34.8% on Wafer, and 1.0% on Freezer.
- Ablation on MIMIC-III: TIMING (zero substitution CPD 0.366) outperforms random-mask IG (RandIG) variants (up to 0.354) and standard IG (0.342).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TIMING instead of IG with segment-based masking | Time-series attribution · MIMIC-III, PAM, Boiler, Epilepsy, Wafer, and Freezer; zero-substitution CPD | CPD: 0.549 -> 0.597; 0.573 -> 0.602; 0.752 -> 1.578; 0.054 -> 0.060; 0.500 -> 0.674; 0.405 -> 0.409 | Helped |
| Use TIMING instead of IG | Time-series attribution · Switch-Feature, State; CPD | CPD: 0.196 -> 0.208; 0.162 -> 0.163 | Mixed |
| Replace independent random point masking with segment-based masking (ablation) | RandIG on MIMIC-III, K=50 · Average and zero substitution | CPD: 0.175 -> 0.177; 0.350 -> 0.366 | Helped |
| Vary RandIG masking probability among p=0.3, p=0.5, and p=0.7 (ablation) | RandIG on MIMIC-III, K=50 · Average and zero substitution | CPD: 0.174; 0.175; 0.175; 0.350; 0.353; 0.354 | No clear effect |
| Vary the TIMING segment parameters (n, s_min, s_max) (ablation) | TIMING on MIMIC-III, K=50 · Average and zero substitution | CPD: 0.173; 0.175; 0.173; 0.176; 0.175; 0.176; 0.175; 0.174; 0.174; 0.177; 0.175 | No clear effect |
Surprising findings
- Conventional IG outperforms recent time series XAI methods (Extrmask, ContraLSP, TimeX++) under the proposed cumulative metrics, and simple ReLU-thresholded IG could appear state-of-the-art under the old simultaneous-masking metrics.
- Masking-based methods such as Extrmask and ContraLSP score well on standard accuracy metrics but get low CPD, suggesting they assign zero importance to negatively important features.
Limitations
- TIMING does not satisfy the completeness axiom of IG, since attributions do not sum to the output difference.
- Synthetic-data results show TIMING does not excel at recovering the true saliency map compared with Extrmask and ContraLSP.
- Gains on some real datasets are small (e.g., 1.0% on Freezer) and the efficient version uses a single sample to approximate the expectation.
Experimental setup
Datasets: Switch-Feature State MIMIC-III PAM Boiler Epilepsy Wafer Freezer
Benchmarks: Switch-Feature State MIMIC-III PAM Boiler Epilepsy Wafer Freezer
Metrics: CPD CPP AUP AUR Accuracy Cross-entropy Sufficiency Comprehensiveness
Baselines: FO AFO GradSHAP DeepLIFT LIME FIT WinIT Dynamask Extrmask ContraLSP TimeX TimeX++
Builds on: Integrated Gradients GradCAM FIT WinIT ContraLSP
Closest papers by meaning
- Temporal Dependencies and Spatio-Temporal Patterns of Time Series Models (AAAI 2024 · similarity 0.86)
- TimeBase: The Power of Minimalism in Efficient Long-term Time Series Forecasting (ICML 2025 · similarity 0.86)
- A Closer Look at Transformers for Time Series Forecasting: Understanding Why They Work and Where They Struggle (ICML 2025 · similarity 0.86)
- InteDisUX: Intepretation-Guided Discriminative User-Centric Explanation for Time Series (AAAI 2025 · similarity 0.86)
- CGS-Mask: Making Time Series Predictions Intuitive for All (AAAI 2024 · similarity 0.85)
- SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (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.