REVIEW 4 major objections 6 minor 19 references
BRATI: Bidirectional Recurrent Attention for Time-Series Imputation
T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash
Pith's one-line read The paper claims that its bidirectional attention-recurrent architecture, BRATI, outperforms existing imputation models across three real-world datasets and three missing-data patterns.
desk verdict Promising architecture and broad experiments, but the Water Quality baseline numbers don't add up and the lack of error bars makes 'consistently outperforms' an overclaim. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the pair of imputation blocks: each block first reweights the embedded, position-encoded input with self-attention, passes it through a GRU, then runs a second attention layer in which the GRU output queries the first attention output, and finally applies a feed-forward network. A sigmoid gate fuses the forward and backward blocks using the input masks and the averaged attention weights from both directions. The training objective sums a masked imputation loss over artificially concealed values, an observed reconstruction loss, and a consistency loss that penalizes disagreement between the two directions. This machinery is what lets the model exploit both temporal order and cross-feature correlations while choosing how much to trust each direction per time step and feature.
What would settle it
Re-train Transformer and SAITS on the Water Quality dataset with longer patience, verified loss convergence, and the same hyperparameter budget, then re-run all six masking configurations; if either baseline's RMSE drops to or below BRATI's in any configuration, the paper's claim of consistent superiority fails.
Extended reading notes
Core claim
The paper claims that a model with two imputation blocks, one reading the series forward and one reading it backward, each pairing multi-head self-attention with a GRU plus a learned attention-weighted fusion of the two directions, gives lower imputation error than M-RNN, BRITS, a Transformer encoder, and SAITS on the PhysioNet 2012, Water Quality, and Space Weather datasets under three missing-data mechanisms. It further introduces the Space Weather dataset as a benchmark with missing-not-at-random gaps caused by instrument saturation. On the paper's own terms, the central discovery is that the bidirectional recurrent attention architecture, trained with masked imputation, observed-reconstruction, and consistency losses, is a stronger and more robust imputer than existing recurrent-only and attention-only designs.
Load-bearing premise
The central comparison stands on the assumption that the baseline models, especially Transformer and SAITS on the Water Quality data, were trained and tuned well enough to represent fair state-of-the-art performance; if those runs failed to converge, the claimed universal superiority is not established.
Editorial extensions
If this is right
- On the three datasets tested, BRATI reports the lowest MAE, RMSE, and MRE in every masking configuration, so its advantage is not limited to random missing values but extends to fixed and variable-length missing runs.
- The ablation results show that removing either the second imputation block or the learned fusion weights degrades accuracy most on variable-length gaps, indicating that both components carry the performance on the hardest scenario.
- The Space Weather benchmark provides a test bed with missingness caused by instrument saturation, which should make missing-not-at-random imputation more accessible to compare.
- Because BRATI uses joint masked-imputation and observed-reconstruction losses plus a consistency loss, it can be trained without requiring data where the missing values are known in advance.
Reading between the lines
- In the paper's Water Quality tables, Transformer and SAITS report RMSE values above 1.3 while their MAE values are around 0.03 to 0.06, a pattern consistent with a few badly predicted points rather than steady error; if those runs did not converge, the comparison there may be unfair and a corrected benchmark could narrow BRATI's reported advantage.
- BRATI's fusion gate and consistency loss could be pulled out and attached to any bidirectional imputer, so the architectural idea may generalize beyond the specific GRU and attention stacking shown here.
- A natural next test is whether BRATI's advantage holds under higher missing rates than 20 percent and on irregularly sampled series, since both conditions are common in real deployments.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes BRATI, a deep-learning architecture for multivariate time-series imputation that combines a bidirectional recurrent GRU layer with multi-head self-attention in two imputation blocks processing the series in opposite temporal directions. The two block representations are combined with learned attention-derived weights, and training uses a joint objective comprising a masked imputation loss, an observed reconstruction loss, and a consistency loss. The authors evaluate BRATI against median imputation, last-value imputation, M-RNN, BRITS, Transformer, and SAITS on three real-world datasets (PhysioNet-2012, Water Quality, and a newly introduced Space Weather dataset) under three synthetic missingness configurations (random MCAR, fixed-length sequences, random-length sequences) at 10% and 20% missing rates, plus ablation experiments. The central claim is that BRATI consistently outperforms all baseline models across all datasets and configurations.
Significance. If the empirical claims held, the contribution would be a solid but incremental architecture that combines well-known components (GRU, multi-head attention, bidirectional processing, and joint-optimization training) and validates them on a broad experimental grid. The paper's strengths are the clear architectural description, the extensive evaluation across 18 dataset-configuration combinations, and the introduction of a new Space Weather dataset that contains naturally MNAR-missing plasma measurements. The ablations (single block vs. full model, learned combination weights vs. simple averaging) are useful and mostly consistent with the design narrative. However, the central 'consistently outperforms' claim is not currently established: the Water Quality tables contain baseline results with implausible MAE/RMSE ratios, no uncertainty quantification is provided for any result, and MNAR performance is claimed but never actually measured. These issues are fixable within the manuscript's scope, but they require substantive re-analysis.
major comments (4)
- [Section 4.5, Tables 4 and 5] The Water Quality MCAR rows for Transformer and SAITS report MAE values of 0.056–0.061 with RMSE values of 1.26–1.45, and the 0.2 series-5 row for SAITS reports MAE 0.0439 with RMSE 0.6124. Such MAE/RMSE ratios are not plausible for converged imputers on standardized data and indicate failed training, a metric artifact, or an evaluation bug. The text itself notes that the self-attention models 'struggle with convergence' in the MCAR scenario, so these rows cannot be treated as valid state-of-the-art comparisons. Because these rows are part of the evidence for the blanket conclusion that BRATI consistently outperforms the baselines, the comparison is not currently supported. Please re-run the affected baselines with convergence diagnostics (validation-loss curves, longer patience, multiple restarts), report what happened, and either present converged results or explicitly exclude non-converged runs.
- [Section 4.5, Tables 2–7] No multi-seed runs, confidence intervals, or statistical significance tests are reported for any experiment. Several reported advantages are very small, for example Space Weather 0.2 range 3-10 MAE 0.0932 vs. 0.0953 and RMSE 0.2291 vs. 0.2317, where the gap is on the order of 1–2%. Without repeated runs or significance testing, the wording 'consistently outperformed' cannot be distinguished from seed noise. Please report means and standard deviations over at least five random seeds, or a paired significance test, for the main comparison tables.
- [Sections 4.1 and 6] The paper repeatedly claims that the Space Weather dataset 'highlight[s] MNAR scenarios' and 'captur[es] complex missing data patterns, including Missing Not At Random (MNAR) scenarios,' but the actual evaluation protocol in Section 4.1 artificially removes observed values uniformly or in sequences and uses those removed values as ground truth. The original MNAR-missing values have no ground truth and are never scored. Thus MNAR performance is claimed but not measured. Either rephrase the contribution to state that the dataset contains natural MNAR missingness while the experiments evaluate synthetic MCAR and sequence-missing patterns, or design an evaluation that scores MNAR positions by holding out known values from instrument-saturation periods where ground truth exists.
- [Section 4.4 and Eq. (24)] The loss weights λMIL, λORL, and λcons are never reported and are not included in the hyperparameter search description, even though they directly determine the objective being optimized. Without these values, the experiments are not reproducible and it is unclear whether the weighting was tuned fairly across models. Please report the chosen weights and their search ranges (or state explicitly that they are fixed constants) for every model.
minor comments (6)
- [Section 3.3.6, Eqs. (18) and (20)] The mask notation is inconsistent: Eq. (18) uses M, Eq. (20) uses \hat M, and the surrounding text refers to 'the input mask M' even though \hat M was introduced earlier as the mask of the artificially masked input. Please clarify which mask is used in each equation.
- [Section 4.4] The sentence 'The dropout is sampled from (0, 0.1, 0.2, 0.3, 0.4, 0.5).' is duplicated in the hyperparameter search description.
- [Section 3.3.5, Eq. (15)] The attention weights γw are described as 'averaged across all the heads,' but no averaging operation is specified in the equation. Please define how the averaged weights are computed from the per-head attention matrices.
- [Section 3.3.7, Eq. (23)] The term 'Discrepancy' is not defined in the paper; earlier Eq. (6) defines the consistency loss via MAE. Please either use MAE explicitly or define Discrepancy.
- [Section 4.1] The sentence 'Experimental results show that the second and third scenarios are considerably more difficult across all the evaluation metrics than the MCAR one, being the last case the most difficult one' states a result before any results are presented; this sentence should be moved to the discussion or rephrased as a preview.
- [Section 5.4] The claim that BRATI 'outperforms all baseline methods in almost every metric in all datasets' is vague; either identify the specific exceptions in Tables 2–7 or change the wording to 'in nearly all configurations.'
Circularity Check
No circular derivation: BRATI's claimed advantage rests on held-out empirical comparison; the only same-author citation affects a data split and is not load-bearing.
full rationale
BRATI is presented as an architecture plus a training loss (Eqs. 13-24); the paper's central claim is that this model scores lower MAE/RMSE/MRE on held-out masked values than baselines. That claim is empirical and self-contained: the test masks are generated independently (Section 4.1), the metrics (Eqs. 25-27) are evaluated only on those artificially masked values, and the training losses (Eqs. 21-24) are not computed on the same masked positions used for final evaluation. No fitted parameter is later renamed as a prediction. The loss weights λMIL, λORL, λcons are mentioned in Eq. 24 as tunable but their values are not reported; this is an under-specification/reproducibility gap, not a constructional equivalence. The only self-citation is [CVMC24] in Section 4.2, used to justify the storm split (20 training / 5 validation / 17 test) for the new Space Weather dataset; that split does not enter the loss, the architecture, or the comparison logic, so it is not load-bearing. The paper itself flags a fairness issue in Section 4.5: "The Self-Attention models exhibit peculiar behavior, struggling with convergence in the MCAR scenario," visible in Tables 4-5 where Transformer/SAITS show MAE ~0.06 with RMSE ~1.3-1.4. That is a baseline-training/correctness concern, not an instance of the paper's result being forced by its own definitions. I therefore find no significant circularity.
Assumptions & free parameters
free parameters (3)
- Consistency factor rho =
0.1
- Loss weights lambda_MIL, lambda_ORL, lambda_cons =
Not reported
- Architecture hyperparameters (N, d_model, d_ffn, heads, dropout) =
Not reported
assumptions (2)
- domain assumption Artificial masking of 10-20% of observed values is a representative proxy for real missing-data performance.
- standard math The attention and GRU equations (Eqs. 9-11) behave as specified in prior literature.
Cite this review
Pith. "Pith review of BRATI: Bidirectional Recurrent Attention for Time-Series Imputation." pith.science (2026). https://pith.science/paper/65CUFC3H
@misc{pith2026250105401,
author = {Pith},
title = {Pith review of: BRATI: Bidirectional Recurrent Attention for Time-Series Imputation},
year = {2026},
howpublished = {\url{https://pith.science/paper/65CUFC3H}},
note = {Machine review of arXiv:2501.05401}
}
read the original abstract
Missing data in time-series analysis poses significant challenges, affecting the reliability of downstream applications. Imputation, the process of estimating missing values, has emerged as a key solution. This paper introduces BRATI, a novel deep-learning model designed to address multivariate time-series imputation by combining Bidirectional Recurrent Networks and Attention mechanisms. BRATI processes temporal dependencies and feature correlations across long and short time horizons, utilizing two imputation blocks that operate in opposite temporal directions. Each block integrates recurrent layers and attention mechanisms to effectively resolve long-term dependencies. We evaluate BRATI on three real-world datasets under diverse missing-data scenarios: randomly missing values, fixed-length missing sequences, and variable-length missing sequences. Our findings demonstrate that BRATI consistently outperforms state-of-the-art models, delivering superior accuracy and robustness in imputing multivariate time-series data.
Figures
Reference graph
Works this paper leans on
-
[9]
CDSA: cross-dimensional self-attention for multivariate, geo-tagged time series imputation
[MSZ+19b] Jiawei Ma, Zheng Shou, Alireza Zareian, Hassan Mansour, Anthony Vetro, and Shih- Fu Chang. CDSA: cross-dimensional self-attention for multivariate, geo-tagged time series imputation. arXiv preprint arXiv:1905.09904 ,
arXiv 1905
-
[14]
Glima: Global and local time series imputation with multi-directional atten- tion learning
[SZX+20] Qiuling Suo, Weida Zhong, Guangxu Xun, Jianhui Sun, Changyou Chen, and Aidong Zhang. Glima: Global and local time series imputation with multi-directional atten- tion learning. In 2020 IEEE International Conference on Big Data (Big Data) , pages 798–807. IEEE,
work page 2020
-
[15]
DeepSD: Supply-Demand Prediction for Online Car-Hailing Services Using Deep Neural Networks
26 [WCLY17] Dong Wang, Wei Cao, Jian Li, and Jieping Ye. DeepSD: Supply-Demand Prediction for Online Car-Hailing Services Using Deep Neural Networks. In 2017 IEEE 33rd International Conference on Data Engineering (ICDE) , pages 243–254,
work page 2017
-
[16]
Deep learning for multivariate time series imputation: A survey
[WDC+24] Jun Wang, Wenjie Du, Wei Cao, Keli Zhang, Wenjia Wang, Yuxuan Liang, and Qingsong Wen. Deep learning for multivariate time series imputation: A survey. arXiv preprint arXiv:2402.04059 ,
-
[18]
Temperature Forecasting via Convolutional Recurrent Neural Networks Based on Time-Series Data
[ZD20] Zao Zhang and Yuan Dong. Temperature Forecasting via Convolutional Recurrent Neural Networks Based on Time-Series Data. Complexity, 2020:1–8, March
work page 2020
-
[19]
ATTAIN: Attention-based Time-Aware LSTM Networks for Disease Progression Modeling
[Zha19] Yuan Zhang. ATTAIN: Attention-based Time-Aware LSTM Networks for Disease Progression Modeling. In In Proceedings of the 28th International Joint Conference on Artificial Intelligence (IJCAI-2019), pp. 4369-4375, Macao, China. ,
work page 2019
-
[1998]
[SLO23] Siyuan Shan, Yang Li, and Junier B. Oliva. Nrtsi: Non-recurrent time series impu- tation. In ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages 1–5,
work page 2023
-
[2006]
Machine learning for predictive maintenance of industrial machines using iot sensor data
[KS17] Ameeth Kanawaday and Aditya Sane. Machine learning for predictive maintenance of industrial machines using iot sensor data. In 2017 8th IEEE International Conference on Software Engineering and Service Science (ICSESS) , pages 87–90,
work page 2017
Show all 19 references
-
[2008]
Uncertainty-aware deep attention recurrent neural network for heterogeneous time series imputation
25 [QID24] Linglong Qian, Zina Ibrahim, and Richard Dobson. Uncertainty-aware deep attention recurrent neural network for heterogeneous time series imputation. arXiv preprint arXiv:2401.02258,
-
[2011]
Missing value im- putation on multidimensional time series
[BDS21] Parikshit Bansal, Prathamesh Deshpande, and Sunita Sarawagi. Missing value im- putation on multidimensional time series. arXiv preprint arXiv:2103.01600 ,
-
[2012]
In 2012 Computing in Cardiology , pages 245–248,
2012
-
[2016]
The arrow of time in mul- tivariate time series
[BSP16] Stefan Bauer, Bernhard Sch¨ olkopf, and Jonas Peters. The arrow of time in mul- tivariate time series. In Maria Florina Balcan and Kilian Q. Weinberger, editors, Proceedings of The 33rd International Conference on Machine Learning , volume 48 of Proceedings of Machine ...
-
[2017]
A survey of missing data imputation using generative adversarial networks
[KTS20] Jaeyoon Kim, Donghyun Tae, and Junhee Seok. A survey of missing data imputation using generative adversarial networks. In 2020 International Conference on Artificial Intelligence in Information and Communication (ICAIIC) . IEEE, February
2020
-
[2018]
ST-MVL: Filling Missing Values in Geo-sensory Time Series Data
[YZZL16] Xiuwen Yi, Yu Zheng, Junbo Zhang, and Tianrui Li. ST-MVL: Filling Missing Values in Geo-sensory Time Series Data. In Proceedings of the 25th International Joint Conference on Artificial Intelligence . IJCAI 2016, June
2016
-
[2019]
Bishift networks for thick cloud removal with multitemporal remote sensing images
[LLJ+23] Chaojun Long, Xinghua Li, Yinghong Jing, Huanfeng Shen, et al. Bishift networks for thick cloud removal with multitemporal remote sensing images. International Journal of Intelligent Systems , 2023,
2023
-
[2020]
Main: Multihead-attention imputation networks
[MPNK21] Spyridon Mouselinos, Kyriakos Polymenakos, Antonis Nikitakis, and Konstantinos Kyriakopoulos. Main: Multihead-attention imputation networks. In 2021 Interna- tional Joint Conference on Neural Networks (IJCNN) , pages 1–8. IEEE,
2021
-
[2021]
Layer normalization
[BKH16] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. arXiv preprint arXiv:1607.06450,
-
[2022]
STING: Self-attention based Time-series Imputation Networks using GAN
[OKJK21] Eunkyu Oh, Taehun Kim, Yunhu Ji, and Sushil Khyalia. STING: Self-attention based Time-series Imputation Networks using GAN. In 2021 IEEE International Conference on Data Mining (ICDM) , pages 1264–1269. IEEE,
2021
-
[2023]
[GAG+13] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. Ch. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley. PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulat...
Reviewed August 10, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.