Pith. sign in

REVIEW 3 major objections 5 minor 44 references

Filling the Missings: Spatiotemporal Data Imputation by Conditional Diffusion

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read CoFILL: a conditional diffusion model that imputes missing spatiotemporal data by fusing temporal and frequency-domain features, outperforming state-of-the-art methods in 12 of 15 configurations.

desk verdict CoFILL is a useful architectural extension of diffusion-based imputation, but the train/test algorithms as written don't match, and the headline gains shouldn't be trusted until the code resolves that. read the letter →

arxiv 2506.07099 v1 pith:JPAARDAO submitted 2025-06-08 cs.LG cs.AI

Wenying He , Jieling Huang , Junhua Gu , Ji Zhang , Yude Bai This is my paper
classification cs.LGcs.AI
keywords spatiotemporalimputationconditionaldiffusionmodelfrequencydomainfeaturescross-attentiontrafficdataairqualitymissingdeeplearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CoFILL is a conditional diffusion model for spatiotemporal data imputation that aims to overcome two weaknesses of prior deep-learning imputers: error accumulation from recursive or autoregressive processing, and incomplete use of the spatial and temporal structure in the data. The paper claims that by conditioning a noise-prediction network on features extracted in parallel from the temporal domain (via temporal convolution and graph convolution) and the frequency domain (via discrete cosine transform), fused by cross-attention, the model can generate missing values that align with the true data distribution. Experiments on air quality (AQI-36) and traffic (METR-LA, PEMS-BAY) datasets under simulated failure, point-missing, and block-missing scenarios show CoFILL achieving the best MAE/MSE in 12 of 15 configurations and a 10.22% MAE reduction over PriSTI on METR-LA Block. The paper argues that avoiding pre-imputation by prior estimates and using a non-recursive diffusion process reduces cumulative error and yields more accurate imputation.

What carries the argument

The key mechanism is a dual-stream conditional feature extraction module that processes the pre-imputed input $X_1$ in parallel through a temporal-domain stream (temporal convolutional network for short-term dynamics and graph convolutional network for spatial dependencies) and a frequency-domain stream (discrete cosine transform for long-term trends and periodic patterns), then fuses these via cross-attention to form the conditioning information $C_{con}$. The noise prediction network $ε_\theta$ then uses this conditioning information through temporal and spatial attention layers during reverse diffusion to transform random noise into meaningful data values. This non-recursive, conditioned diffusion process is what the paper claims avoids cumulative error accumulation and captures both rapid fluctuations and underlying patterns.

What would settle it

A direct falsifier would be to evaluate CoFILL on a real-world spatiotemporal dataset with a reported (non-simulated) missingness pattern that is strongly informative—e.g., missing values caused by systematic sensor drift or correlated outages—and compare its imputation accuracy against the same baselines; if CoFILL's advantage shrinks or reverses, the reliance on random masking would be exposed.

Watch

Extended reading notes

Core claim

The central claim is that a conditional diffusion model that fuses temporal-domain and frequency-domain features through cross-attention can impute missing spatiotemporal data more accurately than existing state-of-the-art methods, without relying on iterative or autoregressive preprocessing. The paper introduces CoFILL, which conditions the reverse diffusion process on a combination of forward-interpolated input and Gaussian-noise-imputed input, and uses a noise prediction network that processes these through temporal (TCN+GCN) and frequency (DCT) streams. The authors report that CoFILL achieves the best performance in 12 out of 15 experimental configurations across three real-world datasets and multiple missing-data patterns, with the largest improvement being a 10.22% MAE reduction over PriSTI on METR-LA Block. They also show through ablations that the forward-interpolation preprocessing and the temporal-domain feature stream are the most critical components, while the cross-attention fusion is the least critical yet still beneficial.

Load-bearing premise

The training-time random masking and the simulated test masks are assumed to represent the missingness mechanism of real deployments, so that the learned conditional distribution transfers to actual data.

Editorial extensions

If this is right

  • If CoFILL's accuracy holds, spatiotemporal imputation in traffic and air-quality monitoring can be performed with lower MAE and MSE than current diffusion-based and deep-learning baselines, improving downstream forecasting and analysis.
  • The non-recursive diffusion design suggests that error accumulation, a known issue in autoregressive and RNN-based imputers, can be mitigated in generative imputation frameworks.
  • The success of fusing frequency-domain features (via DCT) with temporal-domain features indicates that capturing periodic and long-term patterns alongside short-term dynamics is beneficial for imputation accuracy.
  • The ablation results imply that preprocessing with forward interpolation is a key component; removing it degrades performance more than removing either the temporal or frequency stream.
  • The model's performance on different missing-data scenarios (point, block, simulated failure) suggests it generalizes across missingness patterns, though its advantages are uneven (e.g., ties on PEMS-BAY Point).

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's conditioning design implicitly relies on the assumption that the observed portion X1 after forward interpolation provides sufficient information to guide diffusion for the missing part; in cases where missingness is clustered in long contiguous blocks (beyond the tested 4-hour segments), the forward interpolation may become unreliable, potentially limiting performance.
  • The use of DCT over Fourier transform is motivated by avoiding the Gibbs phenomenon, but the paper does not test alternative frequency transforms; similar gains might be obtained with other compact frequency representations, suggesting an empirical question for future work.
  • Although the paper reports improvements on MAE/MSE, CRPS improvements are smaller and on PEMS-BAY Point the CRPS is slightly worse than PriSTI; this hints that the probabilistic calibration of CoFILL is not uniformly superior, which could matter for uncertainty-aware downstream uses.
  • The hyperparameter sensitivity analysis shows that optimal channel size d differs between datasets (16 for AQI-36 vs 64 for traffic), implying that CoFILL's advantage may depend on tuning per domain; a fixed configuration across all spatiotemporal data would likely not reproduce the reported gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript proposes CoFILL, a conditional diffusion model for spatiotemporal data imputation. Preprocessing masks inputs and constructs two initial imputed versions, one by forward interpolation and one by adding Gaussian noise; a dual-stream module extracts temporal and frequency features fused by cross-attention, and a noise-prediction network performs reverse diffusion. Experiments on AQI-36, METR-LA, and PEMS-BAY compare against statistical, matrix/tensor, and deep-learning baselines, with ablations and hyperparameter sensitivity. The main claim is that CoFILL achieves state-of-the-art imputation accuracy, with the best result in 12 of 15 configurations and, for example, a 10.22% MAE reduction over PriSTI on METR-LA Block.

Significance. If the implementation matches the description, the empirical contribution is solid: three real-world datasets, five independent runs, an ablation study, sensitivity analysis, and publicly available source code. The gains over strong diffusion baselines (CSDI, PriSTI) are modest but consistent in most configurations. However, the central claim is currently weakened by an unresolved discrepancy between the training and inference algorithms, and by a contradiction between the abstract's 'without prior estimates' statement and the forward-interpolation conditioning used throughout the method.

major comments (3)
  1. [Section 4.4, Eq. (12), Algorithms 1 and 2] The training and inference procedures are not mutually consistent as written. Algorithm 1 uses the symbol X0 without defining it and does not list the mask M among its inputs, while Eq. (12) trains over the masked-target variable \tilde X_0. Algorithm 2 initializes \tilde X_t ~ N(0,I) at every position and never reinserts observed values during the reverse loop; its input list also omits M even though Section 4.4 states that the observation masks are used. If training noise is applied only to masked positions, then noising all positions at inference is a train/test distribution shift for the denoiser. If X0 is instead the full unmasked tensor, the paper's description of training on imputation targets is misleading and target information leaks into the noisy observation at intermediate steps. Either way, the central empirical claim in Table 2 depends on an algorithm-specification ambiguity that the paper does not resolve. Please define X0 precisely, align the pseudo-code with Eq. (12) and Section 4.4, and state how the mask M is used during reverse sampling (for example, by replacing observed positions with their noised values at each step).
  2. [Abstract, Section 4.1, Table 4] The abstract and introduction claim that CoFILL imputes 'without relying on potentially error-prone prior estimates' and avoids 'error-prone pre-imputation steps,' yet Section 4.1 uses forward interpolation to construct the conditional input X1, and the ablation in Table 4 shows that removing this forward-interpolation condition causes the largest performance degradation. This is an internal inconsistency in the central methodological narrative. The authors should either soften the claim or explicitly distinguish forward interpolation as conditioning information rather than as a prior estimate that the generated values depend on.
  3. [Section 5.5, Tables 2 and 3] The text states that CoFILL shows 'consistent superiority' and 'outperforms state-of-the-art methods,' but this is not true in every reported configuration. On PEMS-BAY Point, Table 2 reports CoFILL MAE 0.57 versus PriSTI 0.55 with identical MSE, and Table 3 reports CRPS 0.0069 versus PriSTI 0.0064. The phrase 'best in 12 of 15 configurations' is accurate, but the broader claim of outperforming SOTA should be qualified to reflect the configuration where CoFILL is not the best.
minor comments (5)
  1. [Section 4.3, last paragraph] The sentence 'The network combines these accumulated features and processes them through two 1-dimensional convolution layers, which contains only the values of the interpolation targets' is grammatically incomplete and unclear; it should specify what the convolution layers output.
  2. [Algorithms 1 and 2] Algorithm 2 uses \tilde X_t ~ N(0,I) as initialization, but the noise schedule and the meaning of \tilde X_t in relation to X1 are not defined; please add a short notation paragraph before the algorithms.
  3. [Section 5.4] The AQI-36 'simulated failure (SF)' pattern is said to replicate real-world sensor malfunction distributions, but no reference or detailed description of the failure model is given; please provide the specific corruption procedure or a citation.
  4. [References and Table 2] There are several presentation issues: 'PEMS-BA Y' in Table 2 should be 'PEMS-BAY', 'BATF[Chen' is missing a space, 'GP-V AE' and 'V AR' have stray spaces, and references [Chen et al., 2022a] and [Chen et al., 2022b] are duplicates of the same paper.
  5. [Section 4.2, Eq. (5)] The symbol T is used both for the number of diffusion steps and for the time length in the DCT formula; please use distinct notation for the time dimension to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the empirical imputation claim is independently testable and no fitted parameter or self-citation carries the central result.

full rationale

The paper's derivation chain is not circular. The model is trained by randomly masking observed values to form imputation targets, extracting conditional features X1 from the masked, forward-interpolated tensor, and minimizing the noise-prediction objective in Eq. (12) over the target positions. At test time, Algorithm 2 runs reverse diffusion to fill the held-out missing masks, and the reported gains are evaluated against independent baselines on SF, Point, and Block patterns. No parameter is fitted to the test targets and then renamed a prediction, and no component of the architecture is defined in terms of the evaluation metric or of the final imputation result. The self-citations present (He et al. 2024, 2025a, 2025b; Bai et al. 2023) support only peripheral claims such as downstream applications, Gaussian-noise enhancement, or a KNN baseline, and they are not load-bearing for the central claim. There is no imported uniqueness theorem and no ansatz smuggled in via citation. The skeptic's concern about Algorithm 1 versus Algorithm 2 training/sampling distribution is a specification and reproducibility risk, not a circularity: an ambiguous X0 and a test-time sampler that may not reinsert observed values would make the method off-distribution, but that is a correctness issue rather than a derivation that reduces to its own inputs. Accordingly, the circularity score is 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central empirical claims rest on standard diffusion training, static graph encoding of spatial relations, temporal-frequency feature complementarity, and the match between training masks and test missingness. No new entities are postulated. The free parameters are conventional network hyperparameters, but they are hand-chosen and, in one case, reported inconsistently.

free parameters (4)
  • Maximum noise level beta_T = 0.2
    Chosen via sensitivity analysis on AQI-36 and METR-LA (Figure 3); controls the noise schedule in the diffusion process.
  • Channel size d = 64 (Table 1), though Section 5.7 says d=16 for AQI-36
    Hidden dimension of feature maps; the paper reports contradictory values, and sensitivity analysis shows it has opposing effects on AQI-36 versus METR-LA.
  • Number of diffusion steps T = 100 for AQI-36; 50 for METR-LA and PEMS-BAY
    Diffusion time horizon set per dataset in Table 1, not derived.
  • Virtual nodes k = 16 for AQI-36; 64 for METR-LA and PEMS-BAY
    Spatial expansion factor in the model; no justification is given beyond the hyperparameter table.
assumptions (4)
  • standard math The standard DDPM corruption and denoising objective L = E||epsilon - epsilon_theta||^2 is a valid training target for conditional imputation.
    Algorithm 1 and eq. (12) assume the conventional diffusion loss without derivation; the reference given (Kingma 2013) is actually the VAE paper rather than DDPM.
  • domain assumption A static adjacency matrix A built from node connections captures the spatial dependencies relevant for imputation.
    Spatial streams (GCN in the conditional module and spatial attention in the noisy estimation module) take A as fixed input in Section 4.2 and eq. (9). If A is misspecified, spatial conditioning is wrong.
  • domain assumption Random masks used during training have a missingness mechanism similar to the test-time masks.
    Training masks are generated by random masking (Section 4.4 and Algorithm 1), while test masks are SF, Point, and Block simulated patterns (Section 5.4). No analysis shows that real missingness matches these distributions.
  • ad hoc to paper Pre-imputation by forward interpolation and the Gaussian-noise version provides useful conditioning rather than biasing the generated values.
    The method relies on X1 produced by forward interpolation (Section 4.1), yet claims not to rely on prior estimates; no theoretical guarantee is provided for this design choice.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Filling the Missings: Spatiotemporal Data Imputation by Conditional Diffusion." pith.science (2026). https://pith.science/paper/JPAARDAO

@misc{pith2026250607099,
  author       = {Pith},
  title        = {Pith review of: Filling the Missings: Spatiotemporal Data Imputation by Conditional Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JPAARDAO}},
  note         = {Machine review of arXiv:2506.07099}
}
read the original abstract

Missing data in spatiotemporal systems presents a significant challenge for modern applications, ranging from environmental monitoring to urban traffic management. The integrity of spatiotemporal data often deteriorates due to hardware malfunctions and software failures in real-world deployments. Current approaches based on machine learning and deep learning struggle to model the intricate interdependencies between spatial and temporal dimensions effectively and, more importantly, suffer from cumulative errors during the data imputation process, which propagate and amplify through iterations. To address these limitations, we propose CoFILL, a novel Conditional Diffusion Model for spatiotemporal data imputation. CoFILL builds on the inherent advantages of diffusion models to generate high-quality imputations without relying on potentially error-prone prior estimates. It incorporates an innovative dual-stream architecture that processes temporal and frequency domain features in parallel. By fusing these complementary features, CoFILL captures both rapid fluctuations and underlying patterns in the data, which enables more robust imputation. The extensive experiments reveal that CoFILL's noise prediction network successfully transforms random noise into meaningful values that align with the true data distribution. The results also show that CoFILL outperforms state-of-the-art methods in imputation accuracy. The source code is publicly available at https://github.com/joyHJL/CoFILL.

Figures

Figures reproduced from arXiv: 2506.07099 by the authors.

Figure 1
Figure 1. Pre-imputing by forward interpolation. measurements, where mi,l = 1 denotes a valid observation at node i and time l, while mi,l = 0 indicates missing data. The spatiotemporal imputation problem aims to reconstruct miss￾ing data in X given the observed measurements and the adja￾cency matrix A. The goal is to develop an imputation method that preserves both the temporal evolution of measurements and their spatial rel… view at source ↗
Figure 2
Figure 2. The framework of CoFILL. where Hin ∈ Rd×N×L, Hˆ in ∈ Rd×N×L, T represents the time step. As m→0, the corresponding DCT component captures low-frequency signals, reflecting long-term stable trends. Conversely, as m→T, it captures high-frequency com￾ponents, indicating periodic structures. Together, the differ￾ent frequency components in DCT represent both the long￾term trends and periodic patterns within the sequence… view at source ↗
Figure 3
Figure 3. Analysis of hyperparameters βT and d [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 37 canonical work pages

  1. [1]

    Traffic flow prediction using graph convolution neural networks

    Anton Agafonov. Traffic flow prediction using graph convolution neural networks. In Proceedings of the 10th international conference on information science and technology (ICIST) , pages 91--95. IEEE, 2020

    work page 2020

  2. [2]

    Missing data imputation using fuzzy-rough methods

    Mehran Amiri and Richard Jensen. Missing data imputation using fuzzy-rough methods. Neurocomputing , 205:152--164, 2016

    work page 2016

  3. [3]

    Learning spatiotemporal latent factors of traffic via regularized tensor factorization: Imputing missing values and forecasting

    Abdelkader Baggag, Sofiane Abbar, Ankit Sharma, Tahar Zanouda, Abdulaziz Al-Homaid, Abhiraj Mohan, and Jaideep Srivastava. Learning spatiotemporal latent factors of traffic via regularized tensor factorization: Imputing missing values and forecasting. IEEE Transactions on Knowledge and Data Engineering (TKDE) , 33(6):2573--2587, 2019

    work page 2019

  4. [4]

    Argusdroid: detecting android malware variants by mining permission-api knowledge graph

    Yude Bai, Sen Chen, Zhenchang Xing, and Xiaohong Li. Argusdroid: detecting android malware variants by mining permission-api knowledge graph. Science China Information Sciences (SCIS) , 66(9):1--19, 2023

    work page 2023

  5. [5]

    E3m: zero-shot spatio-temporal video grounding with expectation-maximization multimodal modulation

    Peijun Bao, Zihao Shao, Wenhan Yang, Boon Poh Ng, and Alex C Kot. E3m: zero-shot spatio-temporal video grounding with expectation-maximization multimodal modulation. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 227--243. Springer, 2025

    work page 2025

  6. [6]

    Crps learning

    Jonathan Berrisch and Florian Ziel. Crps learning. Journal of Econometrics , 237(2):105221, 2023

    work page 2023

  7. [7]

    Brits: Bidirectional recurrent imputation for time series

    Wei Cao, Dong Wang, Jian Li, Hao Zhou, Lei Li, and Yitan Li. Brits: Bidirectional recurrent imputation for time series. Advances in neural information processing systems (NeurIPS) , 31, 2018

    work page 2018

  8. [8]

    Nhits: Neural hierarchical interpolation for time series forecasting

    Cristian Challu, Kin G Olivares, Boris N Oreshkin, Federico Garza Ramirez, Max Mergenthaler Canseco, and Artur Dubrawski. Nhits: Neural hierarchical interpolation for time series forecasting. In Proceedings of the AAAI conference on artificial intelligence (AAAI) , volume 37, pages 6989--6997, 2023

    work page 2023

Show all 44 references
  1. [9]

    Missing traffic data imputation and pattern discovery with a bayesian augmented tensor factorization model

    Xinyu Chen, Zhaocheng He, Yixian Chen, Yuhuan Lu, and Jiawei Wang. Missing traffic data imputation and pattern discovery with a bayesian augmented tensor factorization model. Transportation Research Part C: Emerging Technologies , 104:66--77, 2019

    2019

  2. [10]

    A bayesian tensor decomposition approach for spatiotemporal traffic data imputation

    Xinyu Chen, Zhaocheng He, and Lijun Sun. A bayesian tensor decomposition approach for spatiotemporal traffic data imputation. Transportation research part C: emerging technologies , 98:73--84, 2019

    2019

  3. [11]

    Multitask offloading strategy optimization based on directed acyclic graphs for edge computing

    Jiawen Chen, Yajun Yang, Chenyang Wang, Heng Zhang, Chao Qiu, and Xiaofei Wang. Multitask offloading strategy optimization based on directed acyclic graphs for edge computing. IEEE Internet of Things Journal , 9(12):9367--9378, 2022

    2022

  4. [12]

    Filling the g\_ap\_s: Multivariate time series imputation by graph neural networks

    Andrea Cini, Ivan Marisca, and Cesare Alippi. Filling the g\_ap\_s: Multivariate time series imputation by graph neural networks. arXiv preprint arXiv:2108.00298 , 2021

    2021 arXiv

  5. [13]

    Big learning expectation maximization

    Yulai Cong and Sijia Li. Big learning expectation maximization. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume 38, pages 11669--11677, 2024

    2024

  6. [14]

    Gp-vae: Deep probabilistic time series imputation

    Vincent Fortuin, Dmitry Baranchuk, Gunnar R \"a tsch, and Stephan Mandt. Gp-vae: Deep probabilistic time series imputation. In Proceedings of the International conference on artificial intelligence and statistics (AISTATS) , pages 1651--1661. PMLR, 2020

    2020

  7. [15]

    Generative adversarial networks

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM , 63(11):139--144, 2020

    2020

  8. [16]

    Muse: A deep learning model based on multi-feature fusion for super-enhancer prediction

    Wenying He, Haolu Zhou, Yun Zuo, Yude Bai, and Fei Guo. Muse: A deep learning model based on multi-feature fusion for super-enhancer prediction. Computational Biology and Chemistry , 113:108282, 2024

    2024

  9. [17]

    Xiong, Guangquan Xu, and Fei Guo

    Wenying He, Feiyu Wang, Yude Bai, Neal N. Xiong, Guangquan Xu, and Fei Guo. Pefn: A patches enhancement and hierarchical fusion network for robust vehicle re-identification. IEEE Internet of Things Journal , pages 1--1, 2025

    2025

  10. [18]

    Ensemblese: identification of super-enhancers based on ensemble learning

    Wenying He, Jialu Xu, Yun Zuo, Yude Bai, and Fei Guo. Ensemblese: identification of super-enhancers based on ensemble learning. Briefings in Functional Genomics , 24:elaf003, 04 2025

    2025

  11. [19]

    Fecam: Frequency enhanced channel attention mechanism for time series forecasting

    Maowei Jiang, Pengyu Zeng, Kai Wang, Huan Liu, Wenbo Chen, and Haoran Liu. Fecam: Frequency enhanced channel attention mechanism for time series forecasting. Advanced Engineering Informatics , 58:102158, 2023

    2023

  12. [20]

    Local-global defense against unsupervised adversarial attacks on graphs

    Di Jin, Bingdao Feng, Siqi Guo, Xiaobao Wang, Jianguo Wei, and Zhen Wang. Local-global defense against unsupervised adversarial attacks on graphs. Proceedings of the AAAI Conference on Artificial Intelligence , 37(7):8105--8113, Jun. 2023

    2023

  13. [21]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114 , 2013

    2013 arXiv

  14. [22]

    Missing traffic data: comparison of imputation methods

    Yuebiao Li, Zhiheng Li, and Li Li. Missing traffic data: comparison of imputation methods. IET Intelligent Transport Systems , 8(1):51--57, 2014

    2014

  15. [23]

    Diffusion convolutional recurrent neural network: Data-driven traffic forecasting

    Yaguang Li, Rose Yu, Cyrus Shahabi, and Yan Liu. Diffusion convolutional recurrent neural network: Data-driven traffic forecasting. arXiv preprint arXiv:1707.01926 , 2017

    2017 arXiv

  16. [24]

    Auto-gas: automated proxy discovery for training-free generative architecture search

    Lujun Li, Haosen Sun, Shiwen Li, Peijie Dong, Wenhan Luo, Wei Xue, Qifeng Liu, and Yike Guo. Auto-gas: automated proxy discovery for training-free generative architecture search. In Proceedings of the European Conference on Computer Vision (ECCV) , pages 38--55. Springer, 2025

    2025

  17. [25]

    Gcnet: Graph completion network for incomplete multimodal learning in conversation

    Zheng Lian, Lan Chen, Licai Sun, Bin Liu, and Jianhua Tao. Gcnet: Graph completion network for incomplete multimodal learning in conversation. IEEE Transactions on pattern analysis and machine intelligence (TPAMI) , 45(7):8419--8432, 2023

    2023

  18. [26]

    Swiftsage: A generative agent with fast and slow thinking for complex interactive tasks

    Bill Yuchen Lin, Yicheng Fu, Karina Yang, Faeze Brahman, Shiyu Huang, Chandra Bhagavatula, Prithviraj Ammanabrolu, Yejin Choi, and Xiang Ren. Swiftsage: A generative agent with fast and slow thinking for complex interactive tasks. Advances in Neural Information Processing Syst...

    2024

  19. [27]

    Pristi: A conditional diffusion framework for spatiotemporal imputation

    Mingzhe Liu, Han Huang, Hao Feng, Leilei Sun, Bowen Du, and Yanjie Fu. Pristi: A conditional diffusion framework for spatiotemporal imputation. In Proceedings of the 39th International Conference on Data Engineering (ICDE) , pages 1927--1939. IEEE, 2023

    1927

  20. [28]

    Learning to reconstruct missing data from spatiotemporal graphs with sparse observations

    Ivan Marisca, Andrea Cini, and Cesare Alippi. Learning to reconstruct missing data from spatiotemporal graphs with sparse observations. Advances in Neural Information Processing Systems (NeurIPS) , 35:32069--32082, 2022

    2022

  21. [29]

    Uncertainty-aware variational-recurrent imputation network for clinical time series

    Ahmad Wisnu Mulyadi, Eunji Jun, and Heung-Il Suk. Uncertainty-aware variational-recurrent imputation network for clinical time series. IEEE Transactions on Cybernetics , 52(9):9684--9694, 2021

    2021

  22. [30]

    Enhancing risk prediction in mental health using ensemble hybrid models and administrative healthcare data with irregular intervals

    Faezehsadat Shahidi, M Ethan MacDonald, Dallas Seitz, Rebecca Barry, and Geoffrey Messier. Enhancing risk prediction in mental health using ensemble hybrid models and administrative healthcare data with irregular intervals. In Proceedings of the IEEE International Conference o...

    2024

  23. [31]

    New rnn algorithms for different time-variant matrix inequalities solving under discrete-time framework

    Yang Shi, Chenling Ding, Shuai Li, Bin Li, and Xiaobing Sun. New rnn algorithms for different time-variant matrix inequalities solving under discrete-time framework. IEEE Transactions on Neural Networks and Learning Systems (TNNLS) , 2024

    2024

  24. [32]

    Incomplete multi-view weak-label learning

    Qiaoyu Tan, Guoxian Yu, Carlotta Domeniconi, Jun Wang, and Zili Zhang. Incomplete multi-view weak-label learning. In Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence (IJCAI) , pages 2703--2709, 2018

    2018

  25. [33]

    Csdi: Conditional score-based diffusion models for probabilistic time series imputation

    Yusuke Tashiro, Jiaming Song, Yang Song, and Stefano Ermon. Csdi: Conditional score-based diffusion models for probabilistic time series imputation. Advances in Neural Information Processing Systems (NeurIPS) , 34:24804--24816, 2021

    2021

  26. [34]

    Attention is all you need

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. Attention is all you need. advances in neural information processing systems. Advances in neural information processing systems (NeurIPS) , 30(2017), 2017

    2017

  27. [35]

    Traffic data reconstruction via adaptive spatial-temporal correlations

    Yang Wang, Yong Zhang, Xinglin Piao, Hao Liu, and Ke Zhang. Traffic data reconstruction via adaptive spatial-temporal correlations. IEEE Transactions on Intelligent Transportation Systems (TITS) , 20(4):1531--1543, 2018

    2018

  28. [36]

    Multiple imputation using chained equations: issues and guidance for practice

    Ian R White, Patrick Royston, and Angela M Wood. Multiple imputation using chained equations: issues and guidance for practice. Statistics in medicine , 30(4):377--399, 2011

    2011

  29. [37]

    Graph wavenet for deep spatial-temporal graph modeling

    Zonghan Wu, Shirui Pan, Guodong Long, Jing Jiang, and Chengqi Zhang. Graph wavenet for deep spatial-temporal graph modeling. arXiv preprint arXiv:1906.00121 , 2019

    1906 arXiv

  30. [38]

    Timesnet: Temporal 2d-variation modeling for general time series analysis

    Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. arXiv preprint arXiv:2210.02186 , 2022

    2022 arXiv

  31. [39]

    Ultrahigh thermal stability and piezoelectricity of lead-free knn-based texture piezoceramics

    Lihui Xu, Jinfeng Lin, Yuxuan Yang, Zhihao Zhao, Xiaoming Shi, Guanglong Ge, Jin Qian, Cheng Shi, Guohui Li, Simin Wang, et al. Ultrahigh thermal stability and piezoelectricity of lead-free knn-based texture piezoceramics. Nature Communications , 15(1):9018, 2024

    2024

  32. [40]

    St-mvl: Filling missing values in geo-sensory time series data

    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

    2016

  33. [41]

    Gain: Missing data imputation using generative adversarial nets

    Jinsung Yoon, James Jordon, and Mihaela Schaar. Gain: Missing data imputation using generative adversarial nets. In Proceedings of the International conference on machine learning (ICML) , pages 5689--5698. PMLR, 2018

    2018

  34. [42]

    Temporal regularized matrix factorization for high-dimensional time series prediction

    Hsiang-Fu Yu, Nikhil Rao, and Inderjit S Dhillon. Temporal regularized matrix factorization for high-dimensional time series prediction. Advances in neural information processing systems (NeurIPS) , 29, 2016

    2016

  35. [43]

    Urban computing: concepts, methodologies, and applications

    Yu Zheng, Licia Capra, Ouri Wolfson, and Hai Yang. Urban computing: concepts, methodologies, and applications. ACM Transactions on Intelligent Systems and Technology (TIST) , 5(3):1--55, 2014

    2014

  36. [44]

    write newline

    " write newline "" before.all 'output.state := FUNCTION fin.entry add.period write newline FUNCTION new.block output.state before.all = 'skip after.block 'output.state := if FUNCTION new.sentence output.state after.block = 'skip output.state before.all = 'skip after.sentence '...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.