Research map / Time series / Causal and generative time series
Validation, Robustness, and Accuracy of Perturbation-Based Sensitivity Analysis Methods for Time-Series Deep Learning Models
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 a research plan to benchmark perturbation-based sensitivity analysis methods (Feature Ablation, Feature Occlusion, Morris) across time-series Transformer and linear models on US county COVID-19 case data, evaluated by Spearman rank correlation.
Problem
Interpretability methods for time-series deep learning lack systematic validation, robustness, and accuracy evaluation; it is unclear whether perturbation-based sensitivity analysis methods agree with each other, across models, or with ground truth.
Approach
Benchmark multiple perturbation-group sensitivity analysis methods across several time-series deep learning models (TFT, TimesNet, Autoformer, DLinear, PatchTST) on county-level COVID-19 data, measuring agreement of feature rankings via Spearman correlation and comparing with age-group case ground truth.
Key results
- No quantitative results are reported; the paper describes planned experiments answering three research questions on method agreement, model robustness, and ground-truth alignment.
Limitations
- No experimental results are presented; the work is a proposal of planned experiments.
- Evaluation relies on a single dataset (US county COVID-19 cases) and a single ground-truth source.
- The Spearman correlation with ground truth assumes age-group case counts are a valid importance proxy, which is not validated.
Experimental setup
Datasets: 2-year US county-level COVID-19 cases with 8 population age group features
Benchmarks: US county-level COVID-19 cases
Metrics: Spearman rank correlation coefficient
Baselines: Feature Ablation Feature Occlusion Morris Method
Builds on: Morris method Temporal Fusion Transformer DLinear PatchTST TimesNet
Closest papers by meaning
- Set-Valued Sensitivity Analysis of Deep Neural Networks (AAAI 2025 · similarity 0.82)
- Transformers Learn Low Sensitivity Functions: Investigations and Implications (ICLR 2025 · similarity 0.81)
- SIN: Selective and Interpretable Normalization for Long-Term Time Series Forecasting (ICML 2024 · similarity 0.80)
- Contrasting Adversarial Perturbations: The Space of Harmless Perturbations (AAAI 2025 · similarity 0.79)
- TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation (ICML 2025 · similarity 0.79)
- Towards Trustworthy Deep Learning (AAAI 2024 · similarity 0.79)
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.