Research map / Time series / Causal and generative time series
Learning Perturbations to Explain Time Series Predictions
Joseph Enguehard
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Learn a perturbation generator jointly with the mask instead of using a fixed perturbation | Time-series saliency explanations · HMM | AUR: DynaMask 0.763 (0.026) -> Ours 0.781 (0.013) | Improved |
| Learn a perturbation generator jointly with the mask instead of using fixed perturbations | Time-series saliency explanations · MIMIC-III, 20% masked, average replacement | Cross-Entropy: IG 0.098 (0.006) -> Ours 0.118 (0.008) | Improved |
| Use deletion game instead of preservation game (ablation) | Learned perturbation method · HMM | AUP: 0.885 (0.030) -> 0.346 (0.0034) | Hurt |
| Use deletion game instead of preservation game (ablation) | Learned perturbation method · HMM | AUR: 0.781 (0.013) -> 0.863 (0.012) | Improved |
| Use deletion game instead of preservation game (ablation) | Learned perturbation method · MIMIC-III, zero replacement, 20% masked | Cross-Entropy: 0.318 (0.057) -> 0.079 (0.012) | Hurt |
| Use a unidirectional GRU rather than a Bi-GRU perturbation generator (ablation) | Learned perturbation method · MIMIC-III, 20% masked, average replacement | Cross-Entropy: 0.118 (0.008) -> 0.122 (0.004) | Improved |
| Use a unidirectional GRU rather than a Bi-GRU perturbation generator (ablation) | Learned perturbation method · MIMIC-III, 20% masked, zero replacement | Cross-Entropy: 1.09E-1 (0.023) -> 1.22E-1 (0.008) | Improved |
| Replace the learned perturbation generator with fixed zeros (ablation) | Learned perturbation method · MIMIC-III, 20% masked, zero replacement | Cross-Entropy: 1.22E-1 (0.008) -> 9.64E-2 (0.013) | Hurt |
| Set lambda_1 close to 1 rather than very low or very high (ablation) | Learned perturbation method · HMM | AUP: 0.51 - 0.81; 0.89 - 0.75; 0.99 - 0.26 | Mixed |
| Increase lambda_2 to at least 1 to regularize the perturbation generator (ablation) | Learned perturbation method · HMM | AUP: 0.51 - 0.81 -> 0.89 - 0.75 | Improved |
| Use low lambda_1 values rather than constraining the mask strongly (ablation) | Learned perturbation method · MIMIC-III | Accuracy: 0.926 - 0.348 -> 0.881- 0.534 | Improved |
Closest papers by meaning
- Explaining Time Series via Contrastive and Locally Sparse Perturbations (ICLR 2024 · similarity 0.91)
- Large Pre-trained time series models for cross-domain Time series analysis tasks (NeurIPS 2024 · similarity 0.87)
- Explain Temporal Black-Box Models via Functional Decomposition (ICML 2024 · similarity 0.85)
- Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling (ICML 2024 · similarity 0.85)
- CATS: Enhancing Multivariate Time Series Forecasting by Constructing Auxiliary Time Series as Exogenous Variables (ICML 2024 · similarity 0.85)
- Multivariate Probabilistic Time Series Forecasting with Correlated Errors (NeurIPS 2024 · 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.
