RDDMPI: Residual Denoising Diffusion Model for Probabilistic Multivariate Time Series Imputation
Ramiro Valdes Jara, David Chapman, Adam Meyers
cs.LG · stat.ML
Quick verdict
Residual diffusion over a frozen deterministic imputer claims SOTA on 5 benchmarks; gains are real at high missingness but ablations weaken added mechanisms.
Summary
The paper proposes RDDMPI, a two-stage framework for probabilistic multivariate time series imputation. A pretrained deterministic imputer (T1) first produces a baseline completion and latent representation; a conditional DDPM then operates in residual space, generating corrections only at missing positions. The denoiser conditions on the baseline signal, its latent via FiLM, and a learned reliability gate that adaptively weights baseline information. The authors show residual diffusion is a change of variables of conditional diffusion, with residual second moment decomposing into irreducible uncertainty plus baseline bias, and a Lipschitz bound relating residual-space scores to ideal centered scores. On five benchmarks (ETTh1/2, Weather, Exchange, Illness) under point (0.2-0.8) and block missingness, RDDMPI reports the best MSE/MAE in 18 of 20 aggregated comparisons and lowest CRPS in 8 of 10, beating its own backbone in 19 of 20. Appendix D candidly reports high inference cost (2220 ms/sample on ETTh1 vs 0.33 ms for T1) and a DDIM reduced-step variant with up to +180% MSE degradation on Exchange. Main limitations: gains concentrate at high/block missingness (the full model loses most ablation settings at ratios 0.2-0.6), aggregated tables mask per-ratio losses to T1 on Exchange, comparisons with full-signal diffusion confound architecture differences, and marginal wins lack significance tests.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Reformulating MTSI as a baseline-residual decomposition simplifies the diffusion learning objective, letting it model structured corrections rather than the full missing signal (Sections IV-A, IV-D).
- RDDMPI achieves the best aggregated reconstruction result in 18 of 20 MSE/MAE comparisons and the lowest CRPS in 8 of 10 probabilistic comparisons across five benchmarks (Tables I-II).
- RDDMPI improves over its own deterministic backbone T1 in 19 of 20 aggregated comparisons, with the largest gains at severe missingness (Table IV: up to 39.08% MSE improvement at ETTh1 point-0.8).
- The reliability-aware conditioning mechanism provides its clearest benefit in difficult regimes (point-0.8 and block missingness), per the Exchange ablation (Table III).
Strengths
Honest theoretical framing
Section IV-D explicitly concedes residual diffusion 'corresponds to a conditional reparameterization rather than a fundamentally different inference problem' (Eq 22), then quantifies the practical benefit via the energy decomposition (Eq 24) and a Lipschitz score-perturbation bound (Eq 27) — correct and free of theoretical overclaiming.
Broad, current baselines with per-ratio transparency
Ten baselines including 2024-2026 methods (ImputeFormer KDD'24, FGTI NeurIPS'24, T1 ICLR'26), evaluated at four point ratios plus block missingness over five seeds; Appendices G-H publish full per-ratio means and standard deviations (Tables VIII-XIII).
Candid cost reporting
Regime-honest ablation
Weaknesses
Missing decisive control: data-space diffusion with the same denoiser
Headline components have narrow, single-dataset support
Aggregation strategy flatters results
No statistical testing of marginal wins
Synthetic missingness only; related imputers not compared
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
RDDMPI (arXiv:2609.11648) targets probabilistic multivariate time series imputation (MTSI). The central reframing: split imputation into a pretrained deterministic baseline (T1, ICLR'26) that produces a completed signal plus a latent representation, and a conditional DDPM that diffuses only the missing-region residual R = (1−M)⊙(X−X̃base) (Eq 6), generating correction terms at missing positions. Two conditioning mechanisms are added: a sigmoid reliability gate A = σ(tconv([X̃base, M])) (Eq 17) weighting baseline fusion, and FiLM modulation from the baseline latent (Eqs 19-20). The theory section (IV-D) is unusually candid: Eq 22 establishes residual diffusion is a change of variables of conditional diffusion ('not a fundamentally different inference problem'), Eq 24 decomposes residual energy into irreducible uncertainty plus baseline bias, and Eq 27 bounds the score discrepancy by baseline error under an Lt-Lipschitz assumption. Headline claims: best aggregated MSE/MAE in 18 of 20 comparisons, lowest CRPS in 8 of 10, improvement over its own backbone in 19 of 20.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is RDDMPI and how does it impute missing values in multivariate time series?
RDDMPI is a two-stage probabilistic imputation method. A pretrained deterministic imputer (T1) first fills missing entries to create a baseline completion; a conditional denoising diffusion model (DDPM, 50 steps) then learns to generate only the residual corrections at missing positions, conditioned on the baseline signal, its latent representation, and a learned reliability gate. Sampling 100 residual trajectories yields a predictive distribution; the median serves as the point imputation.
How does residual diffusion differ from standard diffusion imputation like CSDI?
CSDI-style models run diffusion directly on the missing signal starting from pure noise, so the network must learn overall signal structure and uncertainty simultaneously. RDDMPI instead diffuses in residual space around a deterministic baseline, so the network models only baseline errors. The authors show this is formally a change of variables, and argue it is easier to learn when the baseline approximates the conditional mean. RDDMPI reports better MSE/MAE than CSDI in most of 20 aggregated comparisons, though the two methods use different network architectures, so the comparison is not architecture-controlled.
What results does the paper report and on which datasets?
On ETTh1, ETTh2, Weather, Exchange, and Illness (96-length windows) under 20-80% random missingness and block missingness, RDDMPI reports the best MSE/MAE in 18 of 20 aggregated comparisons and lowest CRPS in 8 of 10 against GP-VAE, CSDI, and FGTI, and beats its own T1 backbone in 19 of 20 aggregated comparisons. Gains are largest at 80% missingness and under block missingness; at low missing ratios the deterministic T1 alone can beat the full model (Exchange 0.2: MSE 0.0007 vs 0.0022).
Is code and data available for RDDMPI?
The paper states code is available at github.com/ramirovaldesjara/RDDMPI and uses five public benchmark datasets, with dataset-specific hyperparameters listed in Appendix F. Inference is expensive — roughly 2.2 seconds per sample on ETTh1 for 100 samples at 50 diffusion steps — though a DDIM 10-step variant trades some accuracy for speed, with one large accuracy outlier on Exchange.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are present only as captions in the extracted text; the visual claims in Figure 3 (interval coverage, median tracking) could not be verified from pixel content.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.