pith. sign in

arxiv: 2605.17316 · v1 · pith:FOMGBNXFnew · submitted 2026-05-17 · 💻 cs.LG · cs.AI

Learning Higher-Order Structure from Incomplete Spatiotemporal Data: Multi-Scale Hypergraph Laplacians with Neural Refinement

Keshu Wu , Sixu Li , Zihao Li , Zhiwen Fan , Xiaopeng Li , Yang Zhou This is my paper

Pith reviewed 2026-05-20 14:53 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords multi-scale hypergraphspatiotemporal imputationstructured missingnesshigher-order structuresensor networksresidual refinementgroup conservation patternshypergraph Laplacian
0
0 comments X

The pith

Multi-scale hypergraph Laplacians recover group-conservation patterns in incomplete spatiotemporal data that pairwise graphs cannot access.

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

The paper introduces Multi-Scale Hypergraph Laplacians (MSHL) to learn higher-order structure from sensor data with structured missingness such as clustered outages or whole-sensor blackouts. A discovery stage builds hypergraphs at multiple scales from topology and residual correlations and selects the right scale using only observed values. A refinement stage then applies a hypergraph-conditioned residual network that adds nonlinear corrections only where they help and otherwise defers to the linear estimate. Proofs establish that this captures collective patterns beyond pairwise priors, adapts to the best scale up to a logarithmic factor, and transfers the gain to held-out imputation error. Experiments on real traffic networks show improvement over graph baselines precisely when higher-order structure is present.

Core claim

MSHL is a two-stage framework in which the discovery stage constructs a multi-scale hypergraph from complementary topology and residual-correlation evidence together with an observation-only selector that adapts to the supported interaction scale, while the refinement stage adds a small hypergraph-conditioned residual network that learns nonlinear corrections where informative residual features exist and defers to the linear estimate where they do not. The framework is shown to represent group-conservation patterns inaccessible to pairwise graph priors, to adapt to the best fixed scale up to a logarithmic factor, to transfer this advantage to held-out imputation error, and to admit a one-sid

What carries the argument

Multi-Scale Hypergraph Laplacians (MSHL), a two-stage discovery-plus-refinement framework that builds adaptive multi-scale hypergraphs from observed topology and residuals then applies safe neural corrections.

If this is right

  • MSHL represents group-conservation patterns inaccessible to pairwise graph priors.
  • The method adapts to the best fixed scale up to a logarithmic factor.
  • The scale-adaptation advantage transfers directly to lower held-out imputation error.
  • The refinement stage admits a one-sided guarantee that does not degrade the linear hypergraph estimate.
  • On traffic networks the approach improves over pairwise baselines whenever higher-order structure is identifiable and otherwise matches within sampling noise.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Infrastructure monitoring systems could treat clusters of missing readings as direct signals for discovering group-level conservation laws rather than as isolated values to fill.
  • The observation-only selector mechanism may reduce reliance on manual scale tuning when applying hypergraph models to other domains with incomplete observations.
  • If higher-order structure proves unidentifiable on new data, the framework's fallback behavior ensures performance no worse than standard pairwise methods.
  • The same discovery-plus-safe-refinement pattern could be tested on other spatiotemporal domains such as weather or traffic flow where missingness often occurs in contiguous blocks.

Load-bearing premise

The observation-only selector can reliably identify the appropriate interaction scale from topology and residual-correlation evidence in the presence of structured missingness without access to the missing values themselves.

What would settle it

A collection of spatiotemporal datasets with known higher-order group patterns where the observation-only selector repeatedly selects a scale that yields higher held-out imputation error than the optimal fixed scale, or where MSHL shows no improvement over pairwise-graph baselines under structured missingness beyond sampling noise.

Figures

Figures reproduced from arXiv: 2605.17316 by Keshu Wu, Sixu Li, Xiaopeng Li, Yang Zhou, Zhiwen Fan, Zihao Li.

Figure 1
Figure 1. Figure 1: MSHL overview. Real deployments exhibit diverse missingness from scattered cells to whole-sensor blackouts. MSHL first discovers multi-scale higher-order structure from incomplete observations using complementary topology and residual signals, then refines the linear estimate with a hypergraph-conditioned residual network that defers when no informative residual features are available. The two stages compo… view at source ↗
Figure 2
Figure 2. Figure 2: Multi-scale hypergraph Laplacians. Top row, (a)-(d): a toy network with two planted hyperedges, shown together with the per-scale Laplacians 𝑤𝑠ℒ (𝑠) 𝐻 at scales 𝑠 = 3, 4, 5; each scale contributes the same per-pair regularization energy regardless of 𝑠 (Lemma 1). Bottom row, (e)-(g): on a 60-sensor PEMS￾BAY subnetwork, (e) prior adjacency 𝐴, (f) discovered multi-scale hypergraph ℋb, (g) combined operator 𝐿… view at source ↗
Figure 3
Figure 3. Figure 3: compares MSHL output to ground truth as 𝑁 × 𝑇 heatmaps for all 60 sensors over a 7-day window at 𝑝 = 0.5. On cell-MAR and block-MAR, the imputed heatmaps are visually indistinguishable from ground truth: the diurnal cycle, weekend slowdown, and rush-hour drops are preserved, with no horizontal stripes that would indicate per-sensor bias, no checkerboard texture that would indicate spatial over-smoothing, a… view at source ↗
Figure 4
Figure 4. Figure 4: reports ΔMAE relative to the default for four MSHL hyperparameters on both datasets at 𝑝 = 0.5; [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Component ablation at 𝑝 = 0.5, 𝑁 = 60, 𝑇 = 576. Bars show ΔMAE relative to the default variant: (a) candidate-mechanism ablation, (b) multi-scale vs best single fixed scale, (c) threshold-calibration ablation, (d) shrinkage ablation. Kriging bars near zero throughout reflect deferment [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Numerical signatures of the MSHL theorems. (a) Theorem 2: empirical MSE is affine in 𝜎 2 with 𝑅 2 = 1.0000. (b) Lemma 1: per-pair Dirichlet energy on a constant pattern equals 2/𝑠 to floating-point precision. (c) Theorem 5: per-scale risk versus multi-scale; the selector closely matches the best fixed scale on both datasets. (d) The adaptive-versus-best-fixed gap is well within the Lepski log 𝑆max overhead… view at source ↗
read the original abstract

Sensor networks increasingly govern modern infrastructure, yet the data they lose are rarely missing in the uniform-random patterns assumed by standard imputation benchmarks. Loop detectors go offline during calibration, roadside cabinets silence clusters of nearby sensors, and newly installed instruments provide no history. Such failures create structured absences whose values are constrained by higher-order relations among groups of sensors, not merely by pairwise proximity. Existing low-rank and graph-based methods often miss this collective structure and can fail when missingness becomes coherent. We introduce Multi-Scale Hypergraph Laplacians (MSHL), a two-stage framework for learning higher-order structure from incomplete spatiotemporal observations. The Discovery stage builds a multi-scale hypergraph from complementary topology and residual-correlation evidence, with an observation-only selector that adapts to the supported interaction scale. The Refinement stage adds a small hypergraph-conditioned residual network that is safe by construction: it learns nonlinear corrections where informative residual features exist and defers to the linear estimate where they do not. We prove that MSHL represents group-conservation patterns inaccessible to pairwise graph priors, adapts to the best fixed scale up to a logarithmic factor, transfers this advantage to held-out imputation error, and admits a one-sided refinement guarantee. On two real traffic networks evaluated across scattered cell missingness, contiguous block outages, and whole-sensor blackouts at five rates, MSHL improves over a pairwise-graph baseline whenever higher-order structure is identifiable and otherwise matches it within sampling noise. The results point to a broader principle for reliable infrastructure learning: missing data should be treated not as isolated entries to fill, but as evidence of structure to discover.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes Multi-Scale Hypergraph Laplacians (MSHL), a two-stage framework for imputing incomplete spatiotemporal sensor data under structured missingness. The Discovery stage constructs a multi-scale hypergraph from topology and residual correlations via an observation-only selector that adapts to the supported interaction scale. The Refinement stage adds a hypergraph-conditioned residual network that learns nonlinear corrections where informative and defers to the linear estimate otherwise. The manuscript claims proofs that MSHL represents group-conservation patterns inaccessible to pairwise graphs, adapts to the best fixed scale up to a logarithmic factor, transfers the advantage to held-out imputation error, and admits a one-sided refinement guarantee. Experiments on two traffic networks across scattered, contiguous-block, and whole-sensor missingness at five rates show improvements over pairwise baselines when higher-order structure is present.

Significance. If the selector reliability and theoretical guarantees hold, the work would offer a meaningful advance in handling coherent missingness in infrastructure sensor data by explicitly discovering and exploiting higher-order group structure rather than relying on pairwise graphs or low-rank assumptions. The safety-by-construction property of the residual network and the empirical robustness across multiple missingness patterns are potentially valuable if rigorously supported.

major comments (3)
  1. Abstract and Discovery stage: The adaptation-to-best-scale claim (up to logarithmic factor) and transfer to held-out imputation error rest on the observation-only selector reliably identifying the supported interaction scale from topology and residual correlations alone. No explicit robustness bound or argument is provided showing that residual correlations remain informative about true higher-order groups when missingness is coherent (contiguous blocks or whole-sensor blackouts), which could systematically distort observed residuals and undermine the claimed advantage over pairwise baselines.
  2. Theoretical claims section: The one-sided refinement guarantee and representation of group-conservation patterns appear to invoke external mathematical arguments rather than reducing directly to quantities fitted from the incomplete observations; this introduces a circularity risk for the central claim that MSHL is safe by construction and superior when higher-order structure exists.
  3. §4 (Experiments): The cross-missingness results report gains only when higher-order structure is identifiable, yet no ablation isolates the contribution of the multi-scale selector versus the residual network, making it difficult to confirm that the adaptation mechanism is the load-bearing driver of the reported improvements.
minor comments (2)
  1. Notation for the multi-scale levels and residual network hyperparameters should be introduced with explicit definitions in the main text rather than deferred to the appendix.
  2. Figure captions for the traffic network visualizations could clarify how the selected hyperedges align with the observed residual correlations under each missingness pattern.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive feedback. We address each major comment below with clarifications, proposed additions, and revisions to the manuscript where appropriate.

read point-by-point responses

  1. Referee: Abstract and Discovery stage: The adaptation-to-best-scale claim (up to logarithmic factor) and transfer to held-out imputation error rest on the observation-only selector reliably identifying the supported interaction scale from topology and residual correlations alone. No explicit robustness bound or argument is provided showing that residual correlations remain informative about true higher-order groups when missingness is coherent (contiguous blocks or whole-sensor blackouts), which could systematically distort observed residuals and undermine the claimed advantage over pairwise baselines.

    Authors: We appreciate this observation. The manuscript demonstrates empirical robustness of the selector across contiguous-block and whole-sensor missingness patterns in the traffic datasets, where topology provides a stable anchor and residuals are computed only on observed entries. However, we acknowledge that an explicit robustness argument or bound under coherent missingness is not derived in the current version. In revision we will add a dedicated paragraph in the Discovery stage section that (i) formalizes the conditions under which observed residuals preserve higher-order correlation structure and (ii) includes a short synthetic-data analysis with controlled coherent missingness to illustrate when the selector remains reliable. This addition will make the adaptation claim more rigorously supported without altering the core proofs. revision: partial

  2. Referee: Theoretical claims section: The one-sided refinement guarantee and representation of group-conservation patterns appear to invoke external mathematical arguments rather than reducing directly to quantities fitted from the incomplete observations; this introduces a circularity risk for the central claim that MSHL is safe by construction and superior when higher-order structure exists.

    Authors: We thank the referee for flagging the potential circularity. The proofs are intended to start from the observed residuals and the hypergraph constructed solely from those observations; the group-conservation representation and one-sided guarantee are shown to hold with respect to the linear estimator that would be obtained from the same incomplete data. To eliminate any ambiguity we will revise the theoretical claims section to include a self-contained proof outline that explicitly begins from the observation model and the quantities estimated from incomplete data, thereby removing reliance on external arguments and clarifying the safety-by-construction property. revision: yes

  3. Referee: §4 (Experiments): The cross-missingness results report gains only when higher-order structure is identifiable, yet no ablation isolates the contribution of the multi-scale selector versus the residual network, making it difficult to confirm that the adaptation mechanism is the load-bearing driver of the reported improvements.

    Authors: We agree that an ablation isolating the selector from the refinement network would strengthen the experimental section. In the revised manuscript we will add an ablation study that reports imputation error for (a) the linear baseline, (b) fixed-scale hypergraph without selector, (c) multi-scale selector without refinement, and (d) full MSHL, across all three missingness patterns and rates. This will directly quantify the contribution of the adaptation mechanism. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on independent mathematical arguments

full rationale

The paper's central claims rest on stated proofs that MSHL represents group-conservation patterns inaccessible to pairwise priors, adapts to best scale up to a logarithmic factor, transfers advantage to held-out imputation error, and admits a one-sided refinement guarantee. These are presented as external mathematical results rather than reductions to fitted quantities or self-citations. The Discovery stage's observation-only selector constructs the hypergraph from topology and residual correlations on observed data, but no equation or step equates the selector output directly to the target imputation error by construction. The Refinement stage is described as safe by construction with deferral to linear estimates. No load-bearing self-citation chain or ansatz smuggling is evident that would force the results to the inputs. The method is self-contained against external benchmarks and falsifiable via the reported experiments on real traffic networks.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 2 invented entities

The framework rests on domain assumptions about the existence of extractable higher-order relations and introduces new constructs for multi-scale modeling and safe refinement. Limited free parameters are implied by the adaptive selector and network training.

free parameters (2)
  • multi-scale levels
    Number and choice of scales adapted via observation-only selector from topology and residual correlations
  • residual network hyperparameters
    Size and training of the small hypergraph-conditioned network in the refinement stage
axioms (2)
  • domain assumption Higher-order group relations exist in the sensor data and are identifiable from observed topology and residual correlations
    Invoked to justify hypergraph construction over pairwise graphs
  • ad hoc to paper The observation-only selector can choose the supported interaction scale without access to missing entries
    Central to the Discovery stage adaptation claim
invented entities (2)
  • Multi-Scale Hypergraph Laplacian no independent evidence
    purpose: Model group-conservation patterns at multiple interaction scales
    Core new representation introduced to capture higher-order structure inaccessible to pairwise graphs
  • Hypergraph-conditioned residual network no independent evidence
    purpose: Provide nonlinear corrections only where informative residual features exist
    Introduced to ensure safe refinement that defers to linear estimate otherwise

pith-pipeline@v0.9.0 · 5840 in / 1694 out tokens · 65355 ms · 2026-05-20T14:53:43.053451+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 1 internal anchor

  1. [1]

    Probabilistic recovery of incomplete sensed data in iot.IEEE Internet of Things Journal, 5(4):2282–2292, 2017

    Berihun Fekade, Taras Maksymyuk, Maryan Kyryk, and Minho Jo. Probabilistic recovery of incomplete sensed data in iot.IEEE Internet of Things Journal, 5(4):2282–2292, 2017

    work page 2017

  2. [2]

    Spatial-temporal traffic speed patterns discovery and incompletedatarecoveryviasvd-combinedtensordecomposition.TransportationresearchpartC:emerging technologies, 86:59–77, 2018

    Xinyu Chen, Zhaocheng He, and Jiawei Wang. Spatial-temporal traffic speed patterns discovery and incompletedatarecoveryviasvd-combinedtensordecomposition.TransportationresearchpartC:emerging technologies, 86:59–77, 2018

    work page 2018

  3. [3]

    Reliable sensor deployment for network traffic surveillance

    Xiaopeng Li and Yanfeng Ouyang. Reliable sensor deployment for network traffic surveillance. Transportation research part B: methodological, 45(1):218–231, 2011

    work page 2011

  4. [4]

    Optimization of traffic sensor location for complete link flow observability in traffic network considering sensor failure

    Mostafa Salari, Lina Kattan, William HK Lam, HP Lo, and Mohammad Ansari Esfeh. Optimization of traffic sensor location for complete link flow observability in traffic network considering sensor failure. Transportation Research Part B: Methodological, 121:216–251, 2019

    work page 2019

  5. [5]

    Higher-order interactions shape collective human behaviour

    Federico Battiston, Valerio Capraro, Fariba Karimi, Sune Lehmann, Andrea Bamberg Migliano, Onkar Sadekar, Angel Sánchez, and Matjaž Perc. Higher-order interactions shape collective human behaviour. Nature Human Behaviour, pages 1–17, 2025

    work page 2025

  6. [6]

    Higher-order organization of complex networks

    Austin R Benson, David F Gleich, and Jure Leskovec. Higher-order organization of complex networks. Science, 353(6295):163–166, 2016

    work page 2016

  7. [7]

    Representing higher-order dependencies in networks.Science advances, 2(5):e1600028, 2016

    Jian Xu, Thanuka L Wickramarathne, and Nitesh V Chawla. Representing higher-order dependencies in networks.Science advances, 2(5):e1600028, 2016

    work page 2016

  8. [8]

    A graph-based approach for missing sensor data imputation.IEEE Sensors Journal, 21(20):23133–23144, 2021

    Xiao Jiang, Zean Tian, and Kenli Li. A graph-based approach for missing sensor data imputation.IEEE Sensors Journal, 21(20):23133–23144, 2021

    work page 2021

  9. [9]

    Laplacian convolutional representation for traffic time series imputation.IEEE Transactions on Knowledge and Data Engineering, 36(11):6490–6502, 2024

    Xinyu Chen, Zhanhong Cheng, HanQin Cai, Nicolas Saunier, and Lijun Sun. Laplacian convolutional representation for traffic time series imputation.IEEE Transactions on Knowledge and Data Engineering, 36(11):6490–6502, 2024

    work page 2024

  10. [10]

    The physics of higher-order interactions in complex systems.Nature physics, 17(10):1093–1098, 2021

    Federico Battiston, Enrico Amico, Alain Barrat, Ginestra Bianconi, Guilherme Ferraz de Arruda, Benedetta Franceschiello, Iacopo Iacopini, Sonia Kéfi, Vito Latora, Yamir Moreno, et al. The physics of higher-order interactions in complex systems.Nature physics, 17(10):1093–1098, 2021

    work page 2021

  11. [11]

    The temporal dynamics of group interactions in higher-order social networks.Nature Communications, 15(1):7391, 2024

    Iacopo Iacopini, Márton Karsai, and Alain Barrat. The temporal dynamics of group interactions in higher-order social networks.Nature Communications, 15(1):7391, 2024

    work page 2024

  12. [12]

    Knowledge Is Not Static: Order-Aware Hypergraph RAG for Language Models

    Keshu Wu, Chenchen Kuai, Zihao Li, Jiwan Jiang, Shiyu Shen, Shian Wang, Chan-Wei Hu, Zhengzhong Tu, and Yang Zhou. Knowledge is not static: Order-aware hypergraph rag for language models.arXiv preprint arXiv:2604.12185, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  13. [13]

    Finding the most reliable strategy on stochastic and time- dependent transportation networks: A hypergraph based formulation.Networks and Spatial Economics, 17(3):809–840, 2017

    A Arun Prakash and Karthik K Srinivasan. Finding the most reliable strategy on stochastic and time- dependent transportation networks: A hypergraph based formulation.Networks and Spatial Economics, 17(3):809–840, 2017. 21

    work page 2017

  14. [14]

    Howmuchandwhendoweneedhigher-order information in hypergraphs? a case study on hyperedge prediction

    Se-eunYoon,HyungseokSong,KijungShin,andYungYi. Howmuchandwhendoweneedhigher-order information in hypergraphs? a case study on hyperedge prediction. InProceedings of the web conference 2020, pages 2627–2633, 2020

    work page 2020

  15. [15]

    Equivariant hypergraph diffusion neural operators

    Peihao Wang, Shenghao Yang, Yunyu Liu, Zhangyang Wang, and Pan Li. Equivariant hypergraph diffusion neural operators. InThe Eleventh International Conference on Learning Representations, 2023

    work page 2023

  16. [16]

    Hypergraph-based motion generation with multi-modal interaction relational reasoning.Transportation Research Part C: Emerging Technologies, 180:105349, 2025

    Keshu Wu, Yang Zhou, Haotian Shi, Dominique Lord, Bin Ran, and Xinyue Ye. Hypergraph-based motion generation with multi-modal interaction relational reasoning.Transportation Research Part C: Emerging Technologies, 180:105349, 2025

    work page 2025

  17. [17]

    Ai2-activesafety: Ai-enabled interaction-aware active safety analysis with vehicle dynamics.arXiv preprint arXiv:2505.00322, 2025

    KeshuWu,ZihaoLi,SixuLi,XinyueYe,DominiqueLord,andYangZhou. Ai2-activesafety: Ai-enabled interaction-aware active safety analysis with vehicle dynamics.arXiv preprint arXiv:2505.00322, 2025

    work page arXiv 2025

  18. [18]

    Clustering with hypergraphs: the case for large hyperedges.IEEE transactions on pattern analysis and machine intelligence, 39(9):1697–1711, 2016

    Pulak Purkait, Tat-Jun Chin, Alireza Sadri, and David Suter. Clustering with hypergraphs: the case for large hyperedges.IEEE transactions on pattern analysis and machine intelligence, 39(9):1697–1711, 2016

    work page 2016

  19. [19]

    Multi-hypergraph learning for incomplete multimodality data.IEEE journal of biomedical and health informatics, 22(4):1197–1208, 2017

    Mingxia Liu, Yue Gao, Pew-Thian Yap, and Dinggang Shen. Multi-hypergraph learning for incomplete multimodality data.IEEE journal of biomedical and health informatics, 22(4):1197–1208, 2017

    work page 2017

  20. [20]

    Lau, and Gonzalo R

    Karelia Pena-Pena, Lucas Taipe, Fuli Wang, Daniel L. Lau, and Gonzalo R. Arce. Learning hypergraphs tensor representations from data via t-hgsp.IEEE Transactions on Signal and Information Processing over Networks, 10:17–31, 2024

    work page 2024

  21. [21]

    A survey on hypergraph representation learning.ACM Comput

    Alessia Antelmi, Gennaro Cordasco, Mirko Polato, Vittorio Scarano, Carmine Spagnuolo, and Dingqi Yang. A survey on hypergraph representation learning.ACM Comput. Surv., 56(1), August 2023

    work page 2023

  22. [22]

    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. InInternational Conference on Learning Representations (ICLR ’18), 2018

    work page 2018

  23. [23]

    Tensor completion for estimating missing values in visual data.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):208–220, 2013

    Ji Liu, Przemyslaw Musialski, Peter Wonka, and Jieping Ye. Tensor completion for estimating missing values in visual data.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(1):208–220, 2013

    work page 2013

  24. [24]

    Bayesian temporal factorization for multidimensional time series prediction

    Xinyu Chen and Lijun Sun. Bayesian temporal factorization for multidimensional time series prediction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9):4659–4673, 2022

    work page 2022

  25. [25]

    Low-rank autoregressive tensor completion for spatiotemporal traffic data imputation.IEEE Transactions on Intelligent Transportation Systems, 23(8):12301–12310, 2022

    Xinyu Chen, Mengying Lei, Nicolas Saunier, and Lijun Sun. Low-rank autoregressive tensor completion for spatiotemporal traffic data imputation.IEEE Transactions on Intelligent Transportation Systems, 23(8):12301–12310, 2022

    work page 2022

  26. [26]

    A nonconvex low-rank tensor completion model for spatiotemporal traffic data imputation.Transportation Research Part C: Emerging Technologies, 117:102673, 2020

    Xinyu Chen, Jinming Yang, and Lijun Sun. A nonconvex low-rank tensor completion model for spatiotemporal traffic data imputation.Transportation Research Part C: Emerging Technologies, 117:102673, 2020

    work page 2020

  27. [27]

    Tong Nie, Guoyang Qin, Yunpeng Wang, and Jian Sun. Correlating sparse sensing for large-scale traffic speed estimation: A laplacian-enhanced low-rank tensor kriging approach.Transportation Research Part C: Emerging Technologies, 152:104190, 2023

    work page 2023

  28. [28]

    Filling the g_ap_s: Multivariate time series imputation by graph neural networks.arXiv preprint arXiv:2108.00298, 2021

    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

    work page arXiv 2021

  29. [29]

    Inductive graph neural networks for spatiotemporal kriging.Proceedings of the AAAI Conference on Artificial Intelligence, 35(5):4478–4485, May 2021

    Yuankai Wu, Dingyi Zhuang, Aurelie Labbe, and Lijun Sun. Inductive graph neural networks for spatiotemporal kriging.Proceedings of the AAAI Conference on Artificial Intelligence, 35(5):4478–4485, May 2021

    work page 2021

  30. [30]

    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. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 31. Curran Associates, Inc., 2018

    work page 2018

  31. [31]

    Saits: Self-attention-based imputation for time series.Expert Systems with Applications, 219:119619, 2023

    Wenjie Du, David Côté, and Yan Liu. Saits: Self-attention-based imputation for time series.Expert Systems with Applications, 219:119619, 2023

    work page 2023

  32. [32]

    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. In M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. Wortman Vaughan, editors,Advances in Neural Information Processing Systems, volume 34, pages 24804–24816. Curran Associates, Inc., 2021

    work page 2021

  33. [33]

    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. In2023 IEEE 39th International Conference on Data Engineering (ICDE), pages 1927–1939, 2023. 22

    work page 1927

  34. [34]

    Learningtoreconstructmissingdatafromspatiotemporal graphs with sparse observations

    IvanMarisca,AndreaCini,andCesareAlippi. Learningtoreconstructmissingdatafromspatiotemporal graphs with sparse observations. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 32069–32082. Curran Associates, Inc., 2022

    work page 2022

  35. [35]

    Learning on graph with laplacian regularization.Advances in neural information processing systems, 19, 2006

    Rie Ando and Tong Zhang. Learning on graph with laplacian regularization.Advances in neural information processing systems, 19, 2006

    work page 2006

  36. [36]

    Laplacian regularized low-rank representation and its applications.IEEE transactions on pattern analysis and machine intelligence, 38(3):504–517, 2015

    Ming Yin, Junbin Gao, and Zhouchen Lin. Laplacian regularized low-rank representation and its applications.IEEE transactions on pattern analysis and machine intelligence, 38(3):504–517, 2015

    work page 2015

  37. [37]

    Learning with hypergraphs: Clustering, classification, and embedding

    Dengyong Zhou, Jiayuan Huang, and Bernhard Schölkopf. Learning with hypergraphs: Clustering, classification, and embedding. In B. Schölkopf, J. Platt, and T. Hoffman, editors,Advances in Neural Information Processing Systems, volume 19. MIT Press, 2006

    work page 2006

  38. [38]

    Hypergraph neural networks

    Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural networks. Proceedings of the AAAI Conference on Artificial Intelligence, 33(01):3558–3565, Jul. 2019

    work page 2019

  39. [39]

    Hypergcn: A new method for training graph convolutional networks on hypergraphs

    Naganand Yadati, Madhav Nimishakavi, Prateek Yadav, Vikram Nitin, Anand Louis, and Partha Talukdar. Hypergcn: A new method for training graph convolutional networks on hypergraphs. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 32. Curran Associates, I...

    work page 2019

  40. [40]

    Hypergraphconvolutionandhypergraphattention.Pattern Recognition, 110:107637, 2021

    SongBai,FeihuZhang,andPhilipH.S.Torr. Hypergraphconvolutionandhypergraphattention.Pattern Recognition, 110:107637, 2021

    work page 2021

  41. [41]

    Youareallset: Amultisetfunctionframework for hypergraph neural networks

    EliChien,ChaoPan,JianhaoPeng,andOlgicaMilenkovic. Youareallset: Amultisetfunctionframework for hypergraph neural networks. InInternational Conference on Learning Representations, 2022

    work page 2022

  42. [42]

    Tensor completion for weakly-dependentdataongraphformetropassengerflowprediction.ProceedingsoftheAAAIConference on Artificial Intelligence, 34(04):4804–4810, Apr

    Ziyue Li, Nurettin Dorukhan Sergin, Hao Yan, Chen Zhang, and Fugee Tsung. Tensor completion for weakly-dependentdataongraphformetropassengerflowprediction.ProceedingsoftheAAAIConference on Artificial Intelligence, 34(04):4804–4810, Apr. 2020

    work page 2020

  43. [43]

    On a problem of adaptive estimation in gaussian white noise.Theory of Probability & Its Applications, 35(3):454–466, 1991

    OV Lepskii. On a problem of adaptive estimation in gaussian white noise.Theory of Probability & Its Applications, 35(3):454–466, 1991

    work page 1991

  44. [44]

    A. B. Tsybakov.Introduction to Nonparametric Estimation. Springer, 2009

    work page 2009

  45. [45]

    Greedy function approximation: a gradient boosting machine.Annals of statistics, pages 1189–1232, 2001

    Jerome H Friedman. Greedy function approximation: a gradient boosting machine.Annals of statistics, pages 1189–1232, 2001

    work page 2001

  46. [46]

    Deepresiduallearningforimagerecognition

    KaimingHe,XiangyuZhang,ShaoqingRen,andJianSun. Deepresiduallearningforimagerecognition. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2016

    work page 2016

  47. [47]

    Liqiao Xia, Yongshi Liang, Pai Zheng, and Xiao Huang. Residual-hypergraph convolution network: A model-based and data-driven integrated approach for fault diagnosis in complex equipment.IEEE Transactions on Instrumentation and Measurement, 72:1–11, 2023

    work page 2023

  48. [48]

    Residual enhanced multi-hypergraph neural network

    Jing Huang, Xiaolin Huang, and Jie Yang. Residual enhanced multi-hypergraph neural network. In 2021 IEEE International Conference on Image Processing (ICIP), pages 3657–3661, 2021

    work page 2021

  49. [49]

    Spectrally-normalized margin bounds for neural networks

    Peter Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associates, Inc., 2017

    work page 2017

  50. [50]

    Cambridge university press, 2018

    RomanVershynin.High-dimensionalprobability: Anintroductionwithapplicationsindatascience,volume47. Cambridge university press, 2018. 23 Appendix Theappendixisorganizedasfollows. AppendixAprovidesanotationstableforquickreference. AppendixB contains the IPW lemma and the hypergraph Dirichlet identity. Appendix C gives proofs of all theorems and lemmas in the...

    work page 2018