Research map / Time series / Causal and generative time series
Optimal Information Retention for Time-Series Explanations
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 an information-theoretic Optimal Information Retention principle and the ORTE framework, which learns a binary mask with decoupled straight-through Gumbel-softmax and contrastive learning to produce time-series explanations with low redundancy and high completeness.
Problem
Local explanations of deep time-series classifiers often include redundant patterns or miss key patterns, lacking a unified optimization criterion.
Approach
Casts time-series explanation as conditional mutual information minimization (redundancy) and maximization of completeness, implemented as a learned binary mask trained with contrastive positive/negative pairs, a distribution-alignment term, and an adaptive straight-through estimator (adapt-STE) that uses different forward and backward temperatures.
Key results
- ORTE achieves best performance in 10/12 synthetic cases, improving AUPRC by 7.48% on average over the strongest baseline TIMEX++.
- On ECG, ORTE outperforms TIMEX++ by 8.85% AUPRC, 13.25% AUP and 11.71% AUR.
- Ablation shows adapt-STE improves AUPRC by 7.27% and AUR by 14.22% over removing it on ECG.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ORTE instead of TIMEX++ | Time-series explanation on FreqShapes · FreqShapes | AUPRC: 0.8905±0.0018 -> 0.9998±0.0001 | Helped |
| Use ORTE instead of TIMEX++ | Time-series explanation on SeqComb-UV · SeqComb-UV | AUPRC: 0.8468±0.0014 -> 0.9001±0.0025 | Helped |
| Use ORTE instead of TIMEX++ | Time-series explanation on SeqComb-MV · SeqComb-MV | AUPRC: 0.7589±0.0014 -> 0.8314±0.0019 | Helped |
| Use ORTE instead of TIMEX++ | Time-series explanation on LowVar · LowVar | AUPRC: 0.9466±0.0015 -> 0.9637±0.0016 | Helped |
| Use ORTE instead of TIMEX++ | ECG arrhythmia explanations · ECG | AUPRC: 0.6599±0.0009 -> 0.7183±0.0019 | Helped |
| Add adapt-STE rather than using STE (ablation) | ORTE · ECG | AUPRC: 0.6152±0.0007 -> 0.7183±0.0019 | Helped |
| Add LJS (ablation) | ORTE · ECG | AUPRC: 0.6946±0.0040 -> 0.7183±0.0019 | Helped |
| Add Lct (ablation) | ORTE · ECG | AUPRC: 0.5421±0.0032 -> 0.7183±0.0019 | Helped |
| Add Lcon (ablation) | ORTE · ECG | AUPRC: 0.7130±0.0037 -> 0.7183±0.0019 | Helped |
| Add Lsim (ablation) | ORTE · ECG | AUPRC: 0.7017±0.0040 -> 0.7183±0.0019 | Helped |
| Add Lspr (ablation) | ORTE · ECG | AUPRC: 0.6907±0.0051 -> 0.7183±0.0019 | Helped |
Limitations
- Performance depends on hyperparameter tuning, which can affect instance-level explanations.
- Applying the principle to guide other interpretability methods is left for future research.
Experimental setup
Datasets: FreqShapes SeqComb-UV SeqComb-MV LowVar ECG PAM Epilepsy Boiler
Benchmarks: FreqShapes SeqComb-UV SeqComb-MV LowVar ECG PAM Epilepsy Boiler
Metrics: AUPRC AUP AUR prediction AUROC under occlusion and insertion
Baselines: Integrated Gradients Dynamask WinIT CoRTX SGT + Grad TIMEX TIMEX++
Builds on: TIMEX TIMEX++ Straight-Through Gumbel-Softmax Information Bottleneck CoRTX
Closest papers by meaning
- Explaining Time Series via Contrastive and Locally Sparse Perturbations (ICLR 2024 · similarity 0.87)
- InteDisUX: Intepretation-Guided Discriminative User-Centric Explanation for Time Series (AAAI 2025 · similarity 0.85)
- ReX: A Framework for Incorporating Temporal Information in Model-Agnostic Local Explanation Techniques (AAAI 2025 · similarity 0.85)
- Abstain Mask Retain Core: Time Series Prediction by Adaptive Masking Loss with Representation Consistency (NeurIPS 2025 · similarity 0.85)
- Frequency-Masked Embedding Inference: A Non-Contrastive Approach for Time Series Representation Learning (AAAI 2025 · similarity 0.84)
- Discretizing Continuous Time Series for Imputation with Masked Diffusion Training (NeurIPS 2026 · 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.