Pith. sign in

REVIEW 3 major objections 5 minor 47 references

This paper claims that a diffusion model conditioned on a multi-view user behavior profile outperforms ten baselines for utility data imputation, with up to 29% lower error on block missingness.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 12:12 UTC pith:OV77TPOQ

load-bearing objection Behavior-conditioned diffusion for utility imputation is a genuinely new twist, but the paper as written does not establish that its user profiles come only from training records, and the table-vs-text contradictions make the headline numbers unverifiable. the 3 major comments →

arxiv 2607.29177 v1 pith:OV77TPOQ submitted 2026-07-31 cs.LG cs.AI

MBDiff: Multi-view Behavior-aware Diffusion Model for Probabilistic Utility Data Imputation

Rongchao Xu , Lin Jiang , Dahai Yu , Ximiao Li , Guang Wang This is my paper
classification cs.LG cs.AI
keywords utility data imputationdiffusion modelsuser behavior modelingmulti-view learningmissing datatime series imputationprobabilistic imputationsmart meter data
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tries to establish that missing values in utility time series can be imputed more accurately when the imputation model is guided by the individual user's consumption behavior, not just by the aggregated dataset. To show this, it builds a per-user profile with three views—global day/month/year statistics, latent embeddings of short subsequences, and record-level similarities across dates—and feeds selected parts of that profile into a conditional diffusion denoising network. On three real datasets (electricity, water, gas) with point and block missingness, the method reports lower error than ten baselines, with the largest gains on block missingness: 7.04% better MAE on electricity and 29.1% on water. If the claim holds, it suggests that behavioral context is a reusable lever for improving data quality in utility operations.

Core claim

MBDiff encodes each user's consumption history into a multi-view behavior profile: (i) global patterns—per-day-of-week, per-month, per-year means and variances that give coarse bounds on any data point; (ii) local patterns—latent embeddings of sliding-window subsequences, learned by a Transformer-based variational autoencoder, that capture short-term dependencies; and (iii) instance patterns—record-level latent codes combined with day/month/year matching matrices that identify historically similar dates. The imputation stage selects the most informative references from these views and injects them, along with Z-scores of observed values, into an attentional denoising network that predicts th

What carries the argument

The central object is the multi-view user profile P = (GP, LP, IP), a compact summary of a user's consumption behavior that serves as conditional context for imputation. GP provides statistical bounds, LP provides local temporal shapes, and IP provides record-level similarity; the reference selection module prunes these to the most informative items, and the behavior-aware conditional attentional denoising network (BCADN) fuses them with the noisy record to predict the added noise at each diffusion step.

Load-bearing premise

The evaluation assumes that the user behavior profiles can be built from historical data without leaking test records: the paper splits each user's records 7:1:2 into train/val/test but never states that only training records contribute to the global statistics or to the VAE training; if test records enter the profile, the reported gains are inflated.

What would settle it

Compute the behavior profiles using only the training portion of each user's records (excluding validation and test), retrain MBDiff, and compare point/block imputation MAE on the test set to the numbers in Table 4.2; if the gaps to CSDI and PriSTI shrink or vanish, the leaked-profile hypothesis is confirmed.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Missing utility readings can be filled more accurately when the imputer knows who the user is and how they normally consume, even when historical data are discontinuous.
  • The largest gains appear for block missingness and on sparser datasets (water), which are the cases that most degrade billing and forecasting in practice.
  • The multi-view profile degrades gracefully: at a 50% missing rate on electricity, the method still reports the best point-missing MAE (0.0396) among compared methods.
  • All three views contribute: ablations show that removing any view hurts, with the local-view removal causing the largest drop at severe missingness.
  • The approach has been applied to a real utility provider's data, with over 20 million missing values imputed, indicating practical scalability.

Where Pith is reading between the lines

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

  • If the behavior-profile construction is restricted to training records only (a protocol the paper does not explicitly state), the reported gains might shrink; a re-run with strict train-only profiles would settle whether the improvement is real or partly a leakage artifact.
  • The same three-view conditioning idea could transfer to other user-centric time series—mobility traces, health metrics, appliance-level load—where per-user habits are strong and missingness is structured.
  • The reference-selection scheme is a general efficiency trick: picking a few informative historical examples before diffusion is cheaper than conditioning on the full history, which may make behavior-aware diffusion practical for longer series.
  • A direct testable extension: compare MBDiff on a publicly available smart-meter dataset against CSDI and PriSTI to verify whether the multi-view profile's advantage is dataset-independent.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. This paper proposes MBDiff, a multi-view behavior-aware diffusion model for imputing missing values in user-level utility time series. The method constructs three types of user behavior profiles—global statistics by day/month/year, local subsequence embeddings via a Transformer VAE, and instance-level reference records—and conditions a diffusion denoiser on selected references. The authors evaluate on three utility datasets (electricity, gas, water) from a Florida utility provider, reporting MAE/MSE/CRPS under point and block missingness. The abstract and conclusion claim 7.04% and 29.1% block-missingness improvements on electricity and water, respectively. The paper also includes sensitivity analysis across missing rates, an ablation study, and a brief discussion of related work. The central claim is that behavior-conditioned diffusion imputation outperforms state-of-the-art baselines such as CSDI and PriSTI.

Significance. The idea of combining global, local, and instance-level user behavior as explicit conditioning for a diffusion imputer is well motivated and, if validated, would be a useful practical contribution. The evaluation uses a large real-world dataset (over 120K users), ten baselines, multiple missing rates, and an ablation study, which are strengths. However, the central empirical claims are not verifiable as printed: the behavior-extraction protocol is ambiguous with respect to the train/test split, and several headline numbers in the abstract and §4.2 contradict the tables. No code or data is provided to resolve these ambiguities. The paper has promising content, but it requires substantial revision before its main claims can be accepted.

major comments (3)
  1. [§3.1.1–§3.2.1 vs §4.1.1] The behavior profiles (global, local, and instance) are constructed from D_u, defined in Def. 1 as the user's full historical utility data (Eqs. 3.1–3.12). §4.1.1 states that records are randomly split 7:1:2 into train/validation/test, but it never states that D_u is restricted to the training split. As written, the global means/variances, the TVAE training sets, and the instance-reference pools can include the very test records whose missing values are imputed. This directly affects the paper's core novelty and every reported number. The authors must either explicitly state that only training-split records enter D_u, or rerun the experiments under that protocol and provide code/data to verify. Without this, the claimed gains (e.g., 29.1% water block MAE) cannot be interpreted.
  2. [§4.2, Tables 4.2–4.3] The claim that 'MBDiff achieves the best performance in both point and block missing cases' on electricity is contradicted by the printed table: Table 4.2 reports TimeMixer++ point-missing MAE 0.0199 vs MBDiff 0.0276, so MBDiff is not best on that metric. Additionally, the abstract/conclusion figure of 7.04% improvement for electricity block missingness does not appear anywhere in the tables: from Table 4.2 the block MAE improvement over CSDI is 2.66%, and from Table 4.3 the block CRPS improvement over PriSTI is 7.71%. The statement that MUBE 'ranks in the top two across all settings' is also false as printed: for gas point missing, MBDiff is third behind GP-VAE and PriSTI; for water point missing, it is third behind CSDI and PriSTI. Please reconcile the text with the tables.
  3. [§4.1.3 and Table 4.3] The CRPS comparison is not adequately described. Table 4.3 lists CRPS values for ImputeFormer and TimeMixer++, but these are deterministic methods and the entries are identical to their MAE/MSE values, which suggests a reporting artifact. More importantly, the paper never explains how the predictive distribution is obtained for each probabilistic model (e.g., number of diffusion samples, averaging procedure, or exact CRPS estimator). Since CRPS is one of the headline metrics and is used to support the 'probabilistic imputation' claim, this omission undermines the comparison. Please specify the CRPS estimation protocol and correct Table 4.3.
minor comments (5)
  1. [§3.1.1, Eq. (3.5)] The notation in Eq. (3.5) is ambiguous: 'gdu or gm or gy' should specify which pattern is selected depending on d', m', y'. As written, it looks like an arbitrary choice.
  2. [§4.1.4] The description 'tensors of shape R^{144×48}, representing three days of readings taken at 30-minute intervals' is unclear: the per-record length for electricity is 48, so 144×48 suggests a different grouping. Please clarify.
  3. [§4.1.2] The baseline name 'G-MEAN' in the text appears as 'G-Mean' in Table 4.2. Please standardize names throughout.
  4. [Def. 1] Definition 1 uses N_i both as the number of records for user u and inside D_{u_i}; the subscript is inconsistent.
  5. [§3.2.1] The hyperparameter h_2 (reference count) is never given a value in the implementation details. Please report it.

Circularity Check

0 steps flagged

No significant circularity: the central claim is an empirical comparison against external baselines; the only self-citations are related-work mentions and are not load-bearing.

full rationale

I find no circular step in the paper's derivation chain. MBDiff's claimed contribution is empirical: a diffusion model conditioned on extracted user behavior profiles, evaluated on held-out missing values against external baselines (CSDI, PriSTI, GP-VAE, SAITS, TimeMixer++, etc.). The behavior profiles are learned statistics/reference pools from user history; using them as conditioning is the method's design, not an output that is defined as its input. No equation in Section 3 reduces to a fitted parameter renamed as a prediction, and no uniqueness/ansatz claim is imported from prior work by the authors. The only self-citations (Yu et al. 2026a,b) appear in the related-work survey (Section 5.2) and do not support any load-bearing premise. I note two non-circular concerns: (i) Section 3.1.1 builds global patterns from D_u while Section 4.1.1 splits records 7:1:2 without explicitly stating that D_u is restricted to training records; if test records enter the global statistics, TVAE training, or reference pools, results would be inflated. This is a data-protocol/leakage threat, not a circularity. (ii) Table 4.2 shows TimeMixer++ (0.0199 MAE) beating MBDiff (0.0276) on electricity point missing, contradicting Section 4.2's claim of 'best performance in both point and block missing cases.' These are correctness risks, not derivation-equivalence. Hence circularity score 1.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities. Its assumptions are mostly standard ML background and domain assumptions about behavior stability and missingness. The main concern is the load-bearing data-protocol assumption: behavior profiles may be built from all records including test records, which would inflate the reported gains. The listed free parameters are hyperparameters chosen without sensitivity analysis; none are fitted to produce the headline claim.

free parameters (7)
  • TVAE latent dimension d_z = 64
    Set by hand for behavior embeddings; no sensitivity analysis reported in §4.1.4.
  • sliding window length l_w = 24
    Local-view subsequence length chosen in §4.1.4.
  • sliding window stride s = 12
    Subsequence stride for local-view extraction in §4.1.4.
  • diffusion timesteps T = 500
    DDPM noise schedule length in the second training stage (§4.1.4).
  • noise schedule (β1=1e-4, βT=0.5) = linear
    Chosen in §4.1.4 without sensitivity analysis.
  • reference count h2 = not specified
    Number of selected local/instance references is used in §3.2.1 but never given a value in the text.
  • learning rate / batch size = 1e-3, batch 16
    Training hyperparameters for TVAE and diffusion stages (§4.1.4).
axioms (5)
  • standard math DDPM forward/reverse processes and simplified VLB objective (Ho et al. 2020) are valid and apply to this time-series setting.
    Section 2.2 uses the standard diffusion framework without proof, which is acceptable background.
  • domain assumption The synthetic point/block missing masks used in evaluation resemble real missingness in utility data.
    The paper motivates real-world gaps (residential transitions, sensor failure), but evaluates with random point masks and random-start block masks (§4.2); real missingness may be informative and violate this assumption.
  • domain assumption Past user behavior is stable enough to guide imputation of future target records.
    The entire MUBE design (§3.1) relies on historical utility records being predictive of the target record; behavior changes, occupant changes, or non-stationarity would undermine the reference selection.
  • domain assumption TVAE latent representations preserve the behavior information needed for imputation.
    Equations 3.7 and the local/instance pattern extraction in §3.1.2–3.1.3 assume reconstruction-based VAE latents are a sufficient representation for matching and guiding imputation.
  • domain assumption Global-view and TVAE conditioning statistics are computed without leaking test records into the behavior profile.
    Section 3.1.1 computes global statistics from D_u, the user's full history, while §4.1.1 splits records into train/val/test; the paper does not explicitly restrict behavior extraction to training records, so the evaluation may be optimistic.

pith-pipeline@v1.3.0-daily-deepseek · 17243 in / 14751 out tokens · 130357 ms · 2026-08-03T12:12:26.707644+00:00 · methodology

0 comments
read the original abstract

Utility data (e.g., electricity, water, and gas consumption), collected by ubiquitous sensors and embedded devices, often contains substantial missing values due to various factors such as device failures and data transmission issues. The data missingness can severely impact utility billing accuracy, hinder demand forecasting, and disrupt efficient utility supply management. As a result, utility data imputation has attracted much interest from both industry and academia. While many studies have attempted to address this issue, most of them rely on aggregated datasets for training, overlooking rich user behavior information, which could provide valuable insights for more accurate imputation. However, learning comprehensive user behavior from long-term, diverse, and incomplete utility data remains a significant challenge. Moreover, leveraging user behavior information to guide imputation is nontrivial due to the indirect nature of the correlations. To address these challenges, we propose MBDiff, a Multi-view Behavior-aware Diffusion Model for Probabilistic Utility Data Imputation. MBDiff incorporates two key technical components: (i) a multi-view User Behavior Extraction module that learns comprehensive user behavior from multiple perspectives, including global, local, and instance-level views; and (ii) a behavior-aware conditional diffusion model consisting of a reference selection module and a conditional attentional denoising network to impute utility data in a computationally efficient manner. We implement and evaluate MBDiff by collaborating with one of the largest municipal utility providers in Florida. Experimental results demonstrate our proposed MBDiff effectively outperforms state-of-the-art baselines, e.g., it improves 7.04% and 29.1% on the electricity and water usage datasets for block missingness imputation, respectively.

Figures

Figures reproduced from arXiv: 2607.29177 by Dahai Yu, Guang Wang, Lin Jiang, Rongchao Xu, Ximiao Li.

Figure 3.1
Figure 3.1. Figure 3.1: The pipeline of MBDiff. The shapes outlined by the red and pink lines represent the entire and selected user [PITH_FULL_IMAGE:figures/full_fig_p003_3_1.png] view at source ↗
Figure 3.2
Figure 3.2. Figure 3.2: The architecture of the designed TVAE for utility [PITH_FULL_IMAGE:figures/full_fig_p004_3_2.png] view at source ↗
Figure 3.3
Figure 3.3. Figure 3.3: The architecture of the proposed behavior-aware [PITH_FULL_IMAGE:figures/full_fig_p005_3_3.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

47 extracted references · 7 linked inside Pith

  1. [1]

    Journal of Physics: Conference Series , volume=

    Missing data filling method based on linear interpolation and lightgbm , author=. Journal of Physics: Conference Series , volume=. 2021 , organization=

    2021

  2. [2]

    Applied Mathematics and Computation , volume=

    Robust estimation of traffic density with missing data using an adaptive-R extended Kalman filter , author=. Applied Mathematics and Computation , volume=. 2022 , publisher=

    2022

  3. [3]

    IEEE Transactions on Biomedical Engineering , volume=

    Estimating missing data in temporal data streams using multi-directional recurrent neural networks , author=. IEEE Transactions on Biomedical Engineering , volume=. 2018 , publisher=

    2018

  4. [4]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis , author =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

  5. [5]

    Proceedings of the International Conference on Learning Representations (ICLR) , year =

    Filling the G\_ap\_s: Multivariate Time Series Imputation by Graph Neural Networks , author =. Proceedings of the International Conference on Learning Representations (ICLR) , year =

  6. [6]

    arXiv preprint arXiv:2202.07125 , year=

    Transformers in time series: A survey , author=. arXiv preprint arXiv:2202.07125 , year=

    Pith/arXiv arXiv

  7. [7]

    International conference on artificial intelligence and statistics , pages=

    Gp-vae: Deep probabilistic time series imputation , author=. International conference on artificial intelligence and statistics , pages=. 2020 , organization=

    2020

  8. [8]

    IEEE Transactions on Knowledge and Data Engineering , year=

    Differentiable and scalable generative adversarial models for data imputation , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Csdi: Conditional score-based diffusion models for probabilistic time series imputation , author=. Advances in Neural Information Processing Systems , volume=

  10. [10]

    ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Nrtsi: Non-recurrent time series imputation , author=. ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2023 , organization=

    2023

  11. [11]

    arXiv preprint arXiv:1612.06676 , year=

    Multivariate industrial time series with cyber-attack simulation: Fault detection using an lstm-based predictive data model , author=. arXiv preprint arXiv:1612.06676 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:1905.13628 , year=

    Time series anomaly detection using convolutional neural networks and transfer learning , author=. arXiv preprint arXiv:1905.13628 , year=

    Pith/arXiv arXiv 1905

  13. [13]

    Renewable and Sustainable Energy Reviews , volume=

    Review analysis of COVID-19 impact on electricity demand for residential buildings , author=. Renewable and Sustainable Energy Reviews , volume=. 2021 , publisher=

    2021

  14. [14]

    Applied energy , volume=

    Analytical investigation of autoencoder-based methods for unsupervised anomaly detection in building energy data , author=. Applied energy , volume=. 2018 , publisher=

    2018

  15. [15]

    2017 IEEE Conference on Energy Internet and Energy System Integration (EI2) , pages=

    Short-term electricity load forecasting method based on multilayered self-normalizing GRU network , author=. 2017 IEEE Conference on Energy Internet and Energy System Integration (EI2) , pages=. 2017 , organization=

    2017

  16. [16]

    2018 international joint conference on neural networks (IJCNN) , pages=

    Convolutional neural networks for energy time series forecasting , author=. 2018 international joint conference on neural networks (IJCNN) , pages=. 2018 , organization=

    2018

  17. [17]

    Energy Conversion and Management , volume=

    Smart deep learning based wind speed prediction model using wavelet packet decomposition, convolutional neural network and convolutional long short term memory network , author=. Energy Conversion and Management , volume=. 2018 , publisher=

    2018

  18. [18]

    Neurocomputing , volume=

    Missing data imputation using fuzzy-rough methods , author=. Neurocomputing , volume=. 2016 , publisher=

    2016

  19. [19]

    Journal of Water Resources Planning and Management , volume=

    Dynamic time warping clustering to discover socioeconomic characteristics in smart water meter data , author=. Journal of Water Resources Planning and Management , volume=. 2021 , publisher=

    2021

  20. [20]

    Expert Systems with Applications , volume=

    Saits: Self-attention-based imputation for time series , author=. Expert Systems with Applications , volume=. 2023 , publisher=

    2023

  21. [21]

    2023 IEEE 39th International Conference on Data Engineering (ICDE) , pages=

    Pristi: A conditional diffusion framework for spatiotemporal imputation , author=. 2023 IEEE 39th International Conference on Data Engineering (ICDE) , pages=. 2023 , organization=

    2023

  22. [22]

    arXiv preprint arXiv:2011.11347 , year=

    Time series data imputation: A survey on deep learning approaches , author=. arXiv preprint arXiv:2011.11347 , year=

    Pith/arXiv arXiv 2011

  23. [23]

    2020 , school=

    Deep Learning Based Approaches for Imputation of Time Series Models , author=. 2020 , school=

    2020

  24. [24]

    ACM Computing Surveys , volume=

    Generative adversarial networks in time series: A systematic literature review , author=. ACM Computing Surveys , volume=. 2023 , publisher=

    2023

  25. [25]

    The American Statistician , volume=

    An introduction to kernel and nearest-neighbor nonparametric regression , author=. The American Statistician , volume=. 1992 , publisher=

    1992

  26. [26]

    Advances in neural information processing systems , volume=

    Brits: Bidirectional recurrent imputation for time series , author=. Advances in neural information processing systems , volume=

  27. [27]

    Advances in Neural Information Processing Systems , volume=

    Learning to reconstruct missing data from spatiotemporal graphs with sparse observations , author=. Advances in Neural Information Processing Systems , volume=

  28. [28]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  29. [29]

    Scientific reports , volume=

    Recurrent neural networks for multivariate time series with missing values , author=. Scientific reports , volume=. 2018 , publisher=

    2018

  30. [30]

    International Conference on Machine Learning , pages=

    Probabilistic imputation for time-series classification with missing data , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    2023

  31. [31]

    IEEE Transactions on Cybernetics , volume=

    Uncertainty-aware variational-recurrent imputation network for clinical time series , author=. IEEE Transactions on Cybernetics , volume=. 2021 , publisher=

    2021

  32. [32]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    Generative semi-supervised learning for multivariate time series imputation , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  33. [33]

    arXiv preprint arXiv:2208.09399 , year=

    Diffusion-based time series imputation and forecasting with structured state space models , author=. arXiv preprint arXiv:2208.09399 , year=

    Pith/arXiv arXiv

  34. [34]

    Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    An observed value consistent diffusion model for imputing missing values in multivariate time series , author=. Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  35. [35]

    arXiv preprint arXiv:2402.04059 , year=

    Deep learning for multivariate time series imputation: A survey , author=. arXiv preprint arXiv:2402.04059 , year=

    Pith/arXiv arXiv

  36. [36]

    Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    ImputeFormer: Low rankness-induced transformers for generalizable spatiotemporal imputation , author=. Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  37. [37]

    The Twelfth International Conference on Learning Representations , year=

    Timemixer++: A general time series pattern machine for universal predictive analysis , author=. The Twelfth International Conference on Learning Representations , year=

  38. [38]

    arXiv preprint arXiv:2404.18886 , year=

    A survey on diffusion models for time series and spatio-temporal data , author=. arXiv preprint arXiv:2404.18886 , year=

    arXiv

  39. [39]

    Advances in neural information processing systems , volume=

    Naomi: Non-autoregressive multiresolution sequence imputation , author=. Advances in neural information processing systems , volume=

  40. [40]

    2020 IEEE 6th World Forum on Internet of Things (WF-IoT) , pages=

    From Smart to Intelligent Utility Meters in Natural Gas Distribution Networks , author=. 2020 IEEE 6th World Forum on Internet of Things (WF-IoT) , pages=. 2020 , organization=

    2020

  41. [41]

    IEEE Transactions on smart Grid , volume=

    Review of smart meter data analytics: Applications, methodologies, and challenges , author=. IEEE Transactions on smart Grid , volume=. 2018 , publisher=

    2018

  42. [42]

    Energy and Buildings , volume=

    Heat recovery from urban wastewater: Analysis of the variability of flow rate and temperature , author=. Energy and Buildings , volume=. 2014 , publisher=

    2014

  43. [43]

    Applied Energy , volume=

    A data-driven approach for discovering heat load patterns in district heating , author=. Applied Energy , volume=. 2019 , publisher=

    2019

  44. [44]

    Proceedings of the AAAI Conference on Artificial Intelligence , author=

    TrustEnergy: A Unified Framework for Accurate and Reliable User-level Energy Usage Prediction , volume=. Proceedings of the AAAI Conference on Artificial Intelligence , author=. 2026 , month=. doi:10.1609/aaai.v40i46.41307 , number=

    doi:10.1609/aaai.v40i46.41307 2026

  45. [45]

    arXiv preprint arXiv:2606.00506 , year=

    EnergyMamba: An Uncertainty-Aware Graph-Enhanced Selective State Space Model for Energy Consumption Prediction , author=. arXiv preprint arXiv:2606.00506 , year=

    Pith/arXiv arXiv

  46. [46]

    Proceedings of the 33rd ACM International Conference on Advances in Geographic Information Systems , series=

    TyphoFormer: Language-Augmented Transformer for Accurate Typhoon Track Forecasting , author=. Proceedings of the 33rd ACM International Conference on Advances in Geographic Information Systems , series=. 2025 , publisher=. doi:10.1145/3748636.3763223 , url=

    arXiv 2025

  47. [47]

    Proceedings of the 1st ACM SIGSPATIAL International Workshop on Spatial Intelligence for Smart and Connected Communities , pages=

    Learning from the Storm: A Multivariate Machine Learning Approach to Predicting Hurricane-Induced Economic Losses , author=. Proceedings of the 1st ACM SIGSPATIAL International Workshop on Spatial Intelligence for Smart and Connected Communities , pages=