Learning Where to Look: A Shared Relative-Alignment Module for Time-Series Forecasting and PPG-to-Vital-Sign Reconstruction
Ragamayi Puli, Shunya Nagashima, Shunya Nagashima
cs.LG
Quick verdict
ROOSTER learns a comb bias replacing identity-skip and seasonal-lag conditioning; strong results, but headline wins rest on single-seed, cross-paper evidence.
Summary
The paper proposes ROOSTER, a single conditioning module for two conditional sequence-generation tasks: PPG-to-vital-sign reconstruction (PPG to ECG/respiration) and long-horizon multivariate time-series forecasting. Both tasks traditionally hard-code where each target position reads the condition — an identity skip in reconstruction, seasonal recurrence in forecasting. ROOSTER replaces this with cross-attention from target to condition patches whose logits carry a learned periodic-comb bias over the relative offset, with per-head center, period, and sharpness; the learned period serves as a read-out diagnostic. The module is attached to two host models from the authors' prior work (PENGUIN for reconstruction, DecompSSM for forecasting). On four PPG benchmarks it reports the best heart-rate and respiratory-rate errors versus published baseline numbers (e.g., 9.77 vs 15.64 bpm HR on PPG-DaLiA); a matched control against retrained PENGUIN wins 4 of 5 tasks, with PENGUIN better on BIDMC-RESP. On forecasting it achieves the best horizon-averaged MSE on all four standard datasets (margins of 0.001–0.003 over DecompSSM, single seed) and beats retrained DecompSSM on 20 of 24 dataset–horizon settings under three seeds, while losing on the aperiodic Exchange dataset. Ablations show the relative bias, not content matching, carries the alignment (BIDMC-ECG 1.29 to 6.90 bpm without it). Limitations: cross-paper baseline comparisons, single-seed headline tables, unresolved motion corruption on PPG-DaLiA, and a period read-out that fails on hourly data.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 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
- The identity skip of reconstruction and the seasonal lag of forecasting are two settings of one learnable relative offset, implementable as a periodic-comb bias added to cross-attention logits (Eqs. 1–2).
- ROOSTER outperforms published baselines on all four PPG reconstruction benchmarks under PENGUIN's metric protocol (Table 1), and beats retrained PENGUIN on 4 of 5 tasks under a matched harness (Table 4).
- ROOSTER achieves the best horizon-averaged MSE on all four forecasting datasets (Table 2) and lowers mean MSE versus retrained DecompSSM on 20 of 24 dataset–horizon settings under three seeds, while losing on aperiodic Exchange.
- The relative comb bias, not content matching, carries the alignment: removing it collapses BIDMC-ECG from 1.29 to 6.90 bpm (Table 3).
- The learned per-head center and period are a faithful diagnostic of the data's alignment structure (identity on reconstruction, ~one day on 15-min data, no period on Exchange).
Strengths
Unified conditioning view with a readable diagnostic
The identity skip (δ=0) and seasonal lag (δ=−P) are recast as two settings of one learned offset (Sec. 3.1). The Sec. 4.4 read-out behaves as advertised: P=13.0/11.1 patches ≈ one day on 15-min ETTm1/m2 (12 patches/day), center within 0.1 patches of zero on clean ECG tasks (rediscovering PENGUIN's hard-coded skip), and P=1.7 (no period) exactly on aperiodic Exchange — the same dataset where accuracy degraded.
Decisive ablation isolating the bias
Table 3: removing only the relative bias collapses BIDMC-ECG from 1.29 to 6.90 bpm and PPG-DaLiA from 8.16 to 14.02, while ROOSTER beats the identity skip on all three ECG tasks (1.29 vs 3.06 on BIDMC-ECG) — direct evidence that the bias, not content matching, carries alignment.
Matched-budget controlled evidence
Unusually candid negative results
Well-argued bias-form design
Weaknesses
Headline tables rest on cross-paper, single-seed comparisons
Sub-noise margins in the forecasting headline
Self-referential baseline pool
Auxiliary objective thinly supported
Read-out degrades outside its regime
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
ROOSTER (Secs. 1–3) addresses one question shared by PPG-to-vital-sign reconstruction and long-horizon multivariate forecasting: which condition patch should each target patch read? The central claim is that the identity skip (reconstruction, δ=0) and the seasonal lag (forecasting, δ=−P) are two settings of a single learnable relative offset, realized as a periodic-comb bias b_h(δ)=−κ_h(1−cos(2π(δ−φ_h)/P_h)) added to cross-attention logits (Eqs. 1–2), attached to two host models from the authors' own prior work — PENGUIN (flow matching) for reconstruction and DecompSSM (S5 branches) for forecasting. Three claims are testable: (1) the module improves both hosts; (2) the learned (φ_h, P_h) is a faithful read-out of the alignment structure; (3) the relative bias, not content matching, carries the alignment (Table 3).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 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 ROOSTER and what problem does it solve?
ROOSTER is a conditioning module for models that generate a target sequence from a condition sequence, such as reconstructing an ECG from a wrist PPG signal or forecasting future values of multivariate time series. It adds cross-attention whose scores include a learned periodic 'comb' bias over the offset between target and condition positions, so the model learns where each output position should read from — the same-position alignment used in reconstruction or the one-period-back lag used in seasonal forecasting — instead of hard-coding either choice. The learned period can be read out as an interpretable diagnostic of whether the data contains a usable alignment.
Does ROOSTER beat PENGUIN and DecompSSM on everything?
No. Against a retrained PENGUIN under one shared evaluation harness it wins 4 of 5 reconstruction tasks but loses BIDMC-RESP respiratory reconstruction (1.21 vs 1.56 bpm MAE), and on PPG-DaLiA the margin is near-tie (8.16 vs 8.37 bpm). Against retrained DecompSSM it lowers MSE on 20 of 24 dataset–horizon settings with three seeds, but is 5–11% worse on the aperiodic Exchange dataset at three of four horizons — the dataset where its module found no period.
What periods did the module actually learn?
On 15-minute ETTm1 and ETTm2 electricity/weather data the sharpest attention head learned periods of 13.0 and 11.1 patches; one patch covers two hours, so this is within 10% of one day (12 patches). On PPG-to-ECG reconstruction the learned center settled within 0.1 patches of zero — the same-position alignment that PENGUIN hard-codes as a skip connection. On the aperiodic Exchange dataset it found no usable period (P=1.7), and accuracy degraded there. On hourly ETTh data the read-out failed because the 12-patch lookback spans only three patches per day.
Is code and data available for this paper?
The paper states that its code is publicly available via GitHub (repository Neurogica/ROOSTER). All datasets used are public: PPG-DaLiA, WildPPG, BIDMC, WESAD, and CapnoBase for vital-sign reconstruction; ECL, Weather, the ETT family, Exchange, Solar, and PEMS04 for forecasting. The paper specifies patch length 8, 20k training steps with Adam, and per-head geometric initialization of the period parameter.
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.
- Only Fig. 1 is available (as an embedded caption); the per-horizon forecasting results and the 20/24 harness table are summarized in prose, so individual setting-level numbers cannot be verified from the extracted text.
- Table 1 baselines are cross-paper numbers on different subject splits; the underlying per-baseline protocols could not be checked against the original papers here.
- Some hyperparameters (hidden sizes, number of heads, auxiliary loss weight, exact ensemble size for each task) are not in the extracted text.
- The two prior host papers (PENGUIN arXiv:2602.03858, DecompSSM arXiv:2602.05389) are cited but their content was not independently verified in this review.
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.