Research map / Time series / Causal and generative time series
CGS-Mask: Making Time Series Predictions Intuitive for All
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 CGS-Mask, a model-agnostic perturbation saliency method that uses binary strip masks (consecutive time steps) optimized by a self-adaptive cellular genetic algorithm to explain multivariate time series predictions.
Problem
Existing saliency explanations for time series models score features per time point with fragmented numeric values, ignore temporal ordering, and often require model internals (gradients), making them hard for non-experts to understand.
Approach
Treat consecutive time steps as a cohesive binary strip unit and search for the mask maximizing perturbation error using a cellular genetic algorithm with crossover, mutation, and translation operators, relying only on model inputs and outputs.
Key results
- On synthetic data CGS-Mask reached AUR 0.81 on rare features vs 0.58 for Dynamask and 0.27 for FIT, with mask entropy 0.00 for all binary-mask variants.
- On real-world data CGS-Mask achieved the highest perturbation error on MIMIC-III (0.084 vs 0.072 for Dynamask) and lowest discreteness (46.45 vs 91.88).
- In a pilot user study with 254 participants, over 65% ranked CGS-Mask first and over 85% in their top three; response time 6.26s with 85.4% accuracy vs 19.22s and 40.6% for numeric masks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CGS-Mask instead of Dynamask | time-series saliency explanation · rare feature | AUR: 0.58±0.03 -> 0.81±0.02 | Helped |
| Use CGS-Mask instead of Dynamask | time-series saliency explanation · rare time | AUR: 0.63±0.04 -> 0.82±0.04 | Helped |
| Use CGS-Mask instead of Dynamask | time-series saliency explanation · mixture | AUR: 0.59±0.02 -> 0.82±0.03 | Helped |
| Use CGS-Mask instead of Dynamask | time-series saliency explanation · random | AUR: 0.56±0.03 -> 0.79±0.03 | Helped |
| Use CGS-Mask instead of Dynamask | saliency explanation · MIMIC-III | perturbation error ΔpM: 0.072±0.003 -> 0.084±0.004 | Helped |
| Use CGS-Mask instead of Dynamask | saliency explanation · LSST | perturbation error ΔpM: 1.77±0.09 -> 1.92±0.05 | Helped |
| Use CGS-Mask instead of Dynamask | saliency explanation · NATOPS | perturbation error ΔpM: 1.89±0.06 -> 2.15±0.09 | Helped |
| Use CGS-Mask instead of Dynamask | saliency explanation · AE | perturbation error ΔpM: 121.15±2.93 -> 131.24±3.72 | Helped |
| add temporally contiguous strips rather than point-based masks (ablation) | CGS-Mask · synthetic data sets | mask discreteness DM | Helped |
| add the cellular genetic algorithm rather than standard genetic algorithm (ablation) | CGS-Mask · real-world data sets | perturbation error ΔpM: 0.080±0.003 -> 0.084±0.004; 1.87±0.05 -> 1.92±0.05; 1.99±0.05 -> 2.15±0.09; 122.72±4.69 -> 131.24±3.72 | Helped |
| add the translation operator (ablation) | CGS-Mask · real-world data sets | perturbation error ΔpM: 0.082±0.003 -> 0.084±0.004; 1.89±0.04 -> 1.92±0.05; 2.09±0.07 -> 2.15±0.09; 127.56±3.54 -> 131.24±3.72 | Helped |
Surprising findings
- Removing the strip component (w/o Strip) gave the highest AUR in some synthetic settings and only modestly lower perturbation error, while the strip mainly drove mask continuity.
Limitations
- Focuses on local interpretability only; global explanations not supported.
- Real-world evaluation relies on perturbation error since no ground-truth explanations exist.
- User study is a pilot survey with questionnaire-based, not controlled, evaluation.
Experimental setup
Datasets: MIMIC-III LSST NATOPS Appliances Energy synthetic rare features synthetic rare time synthetic mixture synthetic random
Benchmarks: MIMIC-III LSST NATOPS Appliances Energy synthetic rare feature synthetic rare time synthetic mixture synthetic random
Metrics: AUP AUR mask discreteness mask entropy perturbation error (cross entropy / squared error) user preference rank response time and accuracy
Baselines: Feature Occlusion Feature Permutation Integrated Gradient Shapley Value Sampling FIT RISE DeepLIFT Dynamask
Builds on: Dynamask FIT Integrated Gradients RISE LIME
Closest papers by meaning
- Abstain Mask Retain Core: Time Series Prediction by Adaptive Masking Loss with Representation Consistency (NeurIPS 2025 · similarity 0.85)
- TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation (ICML 2025 · similarity 0.85)
- Language Pre-training Guided Masking Representation Learning for Time Series Classification (AAAI 2025 · similarity 0.84)
- TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting (NeurIPS 2025 · similarity 0.83)
- Discretizing Continuous Time Series for Imputation with Masked Diffusion Training (NeurIPS 2026 · similarity 0.83)
- SynTSBench: Rethinking Temporal Pattern Learning in Deep Learning Models for Time Series (NeurIPS 2025 · similarity 0.83)
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.