Research map / Time series / Causal and generative time series
Towards Editing Time Series
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.
Introduces Time Series Editing (TSE), editing attributes of an existing time series, and proposes TEdit, a multi-resolution diffusion model trained with a bootstrap self-scoring learning algorithm.
Problem
Existing time series generation is either unconditional or conditional and generates from scratch, so it cannot modify attributes of an existing sample while preserving other properties (sample-level "what if" editing). Data coverage over attribute combinations is sparse, and attributes act at diffe
Approach
Edit a time series via a diffusion model by DDIM-inverting the source series under source attributes and regenerating under target attributes; add a multi-resolution patch-based noise estimator (with parallel attention masking) and a bootstrap finetuning step that self-scores generated edits by round-trip editing MSE and retrains on the top-scoring samples.
Key results
- On the Synthetic dataset, TEdit-CSDI reaches MSE 0.1235 and MAE 0.2606 versus 0.1789 and 0.3221 for CSDI.
- TEdit-TW improves edited-attribute RaTS on Synthetic to 1.0121 from Time Weaver's 0.9030, and CTAP to 0.7957 from 0.6943.
- Editing reduces MSE vs conditional generation on Synthetic (0.1235 vs 0.2581 for TEdit-CSDI), with conditional generation producing samples near the dataset mean.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add TEdit editing with multi-resolution modeling and bootstrap learning instead of conditional generation baseline Time Weaver | Time Weaver · Synthetic | MSE: 0.1454 -> 0.1315 | Helped |
| Add TEdit editing with multi-resolution modeling and bootstrap learning instead of Time Weaver | Time Weaver · Air Quality | CTAP (edited attributes): 0.3266 -> 0.3930 | Helped |
| Add TEdit editing with multi-resolution modeling and bootstrap learning instead of Time Weaver | Time Weaver · Motor Imagery | CTAP (edited attributes): 0.4168 -> 0.4348 | Helped |
| Add bootstrap learning (ablation) | TEdit-TW · Synthetic | MSE: 0.1376 -> 0.1315 | Helped |
| Add bootstrap learning (ablation) | TEdit-TW · Air Quality | CTAP (edited attributes): 0.3792 -> 0.3930 | Helped |
| Add bootstrap learning (ablation) | TEdit-TW · Air Quality | CTAP (preserved attributes): 0.6418 -> 0.6274 | Hurt |
| Add bootstrap learning (ablation) | TEdit-TW · Motor Imagery | CTAP (edited attributes): 0.4289 -> 0.4348 | Helped |
| Add bootstrap learning (ablation) | TEdit-TW · Motor Imagery | CTAP (preserved attributes): 0.6658 -> 0.6621 | No clear effect |
| Add multi-resolution modeling (ablation) | TEdit-TW · Synthetic | CTAP (edited attributes): 0.6943 -> 0.7957 | Helped |
| Add multi-resolution modeling (ablation) | TEdit-TW · Air Quality | CTAP (edited attributes): 0.3266 -> 0.3930 | Helped |
| Add multi-resolution modeling (ablation) | TEdit-TW · Motor Imagery | CTAP (edited attributes): 0.4168 -> 0.4348 | Helped |
| Use ground-truth source-target pairs for fine-tuning instead of bootstrap samples (ablation) | TEdit-TW · Synthetic | MSE: 0.1315 -> 0.1233 | Helped |
Surprising findings
- Bootstrap learning with self-scored pseudo-targets nearly matches training with ground-truth target series on attribute-level metrics, though GT is better on overall MSE/MAE.
- On the Air Quality dataset, bootstrapping improves edited-attribute scores at the cost of preserved-attribute scores, suggesting entanglement between city and season.
Limitations
- Different attributes vary in editing difficulty, and some attributes are hard to edit without affecting others.
- Experiments limited to categorical attributes and a single PM2.5 signal/one EEG setup; generalization is not broadly tested.
Experimental setup
Datasets: Synthetic (generated) Air Quality (KDD Cup 2018) Motor Imagery (BCI EEG)
Benchmarks: Synthetic TSE benchmark Air Quality Motor Imagery
Metrics: MSE MAE RaTS |RaTS| CTAP score
Baselines: CSDI Time Weaver Conditional generation (no editing)
Builds on: CSDI Time Weaver DDIM CLIP Patch-based Transformer (PatchTST)
Closest papers by meaning
- VerbalTS: Generating Time Series from Texts (ICML 2025 · similarity 0.85)
- DeformableTST: Transformer for Time Series Forecasting without Over-reliance on Patching (NeurIPS 2024 · similarity 0.84)
- TimeEmb: A Lightweight Static-Dynamic Disentanglement Framework for Time Series Forecasting (NeurIPS 2025 · similarity 0.84)
- TimeDP: Learning to Generate Multi-Domain Time Series with Domain Prompts (AAAI 2025 · similarity 0.83)
- Explaining Time Series via Contrastive and Locally Sparse Perturbations (ICLR 2024 · similarity 0.83)
- Time Series Generation Under Data Scarcity: A Unified Generative Modeling Approach (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.