Research map / Time series / Causal and generative time series
Self-attention-based Diffusion Model for Time-series Imputation in Partial Blackout Scenarios
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 SADI, a self-attention-based conditional diffusion model with a two-stage, learned-weight imputation process, and a "partial blackout" missingness framework for multivariate time-series imputation.
Problem
Imputing multivariate time series under partial blackouts, where subsets of features are missing for consecutive time steps, a more general missingness pattern than random, complete blackout, or forecasting cases.
Approach
A single-channel conditional diffusion denoiser combining a Feature Dependency Encoder (dilated 1-D convolution plus self-attention over features), Gated Temporal Attention blocks (self-attention with GLU, DiffWave/WaveNet-style residuals), a second refinement stage, and a learned sigmoid weighting of the two stage outputs; trained with a mixed partial-blackout fine-tuning strategy.
Key results
- On Electricity with 1 missing feature, SADI-MPB reaches MSE 1.70e-01 vs CSDI 3.56e-01 and SAITS 8.31e-01 (MSE, 30-step blocks).
- On AgAID with 1 missing feature, SADI-MPB CRPS 1.09e-03 vs CSDI 7.26e-03 and SADI-RM 1.79e-03.
- Ablation on NACSE (10 features missing): removing FDE raises MSE from 4.88e-03 to 3.74e-02, showing feature dependency modeling matters most on high-correlation datasets.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune SADI with mixed partial-blackout and random-missing training instead of random-missing training alone | SADI · Partial-blackout imputation across AgAID, Air Quality, Electricity, and NACSE | MSE: 5.98e-04 $\pm$ 2.29e-04 -> 2.93e-04 $\pm$ 1.04e-04 (AgAID, 3 missing features) | Helped |
| Add FDE (ablation) | SADI · Partial-blackout imputation across AgAID, Air Quality, Electricity, and NACSE | MSE: 1.43e-03 $\pm$ 9.00e-04 -> 2.93e-04 $\pm$ 1.04e-04 (AgAID, 3 missing features); 3.74e-02 $\pm$ 2.04e-02 -> 4.88e-03 $\pm | Helped |
| Add second-stage imputation (ablation) | SADI · Partial-blackout imputation across AgAID, Air Quality, Electricity, and NACSE | MSE: 1.12e-03 $\pm$ 3.10e-04 -> 2.93e-04 $\pm$ 1.04e-04 (AgAID, 3 missing features); 4.10e-02 $\pm$ 2.58e-02 -> 4.88e-03 $\pm | Helped |
| Add learned weighted combination of the two intermediate imputations (ablation) | SADI · Partial-blackout imputation across AgAID, Air Quality, Electricity, and NACSE | MSE: 7.81e-04 $\pm$ 5.62e-04 -> 2.93e-04 $\pm$ 1.04e-04 (AgAID, 3 missing features); 3.95e-02 $\pm$ 1.61e-02 -> 4.88e-03 $\pm | Helped |
Surprising findings
- Second-stage refinement can sometimes degrade accuracy, motivating the learned weighted combination.
- CSDI had to be reduced to 8 channels on Electricity and NACSE due to GPU memory limits, possibly hurting its reported performance.
Limitations
- CSDI had to be run with reduced channels (8) on high-dimensional datasets, which may bias comparison.
- Second-stage imputation can degrade accuracy, requiring a learned weighting to compensate.
- Synthetic-data results are only in supplementary material and benefit depends on feature interdependency.
Experimental setup
Datasets: AgAID Air Quality Electricity NACSE
Benchmarks: AgAID Air Quality Electricity NACSE
Metrics: MSE CRPS
Baselines: CSDI BRITS SAITS MICE
Builds on: CSDI SAITS DiffWave WaveNet
Closest papers by meaning
- Discretizing Continuous Time Series for Imputation with Masked Diffusion Training (NeurIPS 2026 · similarity 0.87)
- Diffusion Transformers for Imputation: Statistical Efficiency and Uncertainty Quantification (NeurIPS 2025 · similarity 0.87)
- RDPI: A Refine Diffusion Probability Generation Method for Spatiotemporal Data Imputation (AAAI 2025 · similarity 0.86)
- ProCTI: Prototype-Refined Global Conditioning for Diffusion-Based Time Series Imputation (NeurIPS 2026 · similarity 0.85)
- Frequency-aware Generative Models for Multivariate Time Series Imputation (NeurIPS 2024 · similarity 0.85)
- BayOTIDE: Bayesian Online Multivariate Time Series Imputation with Functional Decomposition (ICML 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.