Pith. sign in

REVIEW 4 major objections 7 minor 52 references

User Trajectory Prediction Unifying Global and Local Temporal Information

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that concurrently extracting global patterns with an MLP and multi-scale local patterns with a CNN, then fusing them with cross-attention, reduces user trajectory prediction error by about five percent relative to the…

desk verdict A plausible new architecture with careful experiments, but missing error bars and a hand-wavy theory section; worth reviewing. read the letter →

arxiv 2508.02161 v1 pith:N2UN7NYI submitted 2025-08-04 cs.LG

Wei Hao , Bin Chong , Ronghua Ji , Chen Hou This is my paper · ORCID
classification cs.LG
keywords usertrajectorypredictionglobaltemporalinformationlocalmulti-scaleconvolutionalneuralnetworkcross-attentionMLPGeoLifedatasettimeseriesforecasting
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

This paper tries to establish that user trajectory prediction improves when a model captures global temporal patterns and multi-scale local patterns at the same time, rather than relying on one or extracting them sequentially. It introduces MMCTP, built from an MLP branch, a multi-scale CNN branch, and a cross-attention fusion block, and reports that this combination gives the lowest mean squared error and mean absolute error among compared models on the GeoLife dataset. In the headline configuration, MMCTP reduces MSE by 5.04% and MAE by 4.35% against ModernTCN at 12-step prediction while keeping inference time similar. This matters because more accurate trajectory forecasts can drive proactive edge service migration and other mobility services, and the architecture remains linear-time in sequence length.

What carries the argument

The central object is the MMCTP architecture: an MLP branch with a global receptive field; an MSCNN branch that takes the most recent $R=24$ time steps concatenated with zero-padded future positions and applies 1D convolutions with kernel sizes 3, 5, and 7; and a multi-head cross-attention block that uses local features as queries and global features as keys and values. The MLP branch also uses RevIN normalization and positional, temporal, and value embeddings before the linear layers. The architecture's claimed time complexity is $O(q)$ in sequence length $q$, matching linear models while adding multi-scale local information.

What would settle it

Re-run the 12-step prediction comparison on the 15s GeoLife split with at least ten seeds and report the standard deviation of MSE for MMCTP and ModernTCN; if the 5.04% gap is smaller than one standard deviation, the claimed ordering is not established.

Watch

Extended reading notes

Core claim

The central discovery, on the paper's own terms, is that global and local temporal information are complementary for trajectory prediction and are best extracted in parallel and then fused with cross-attention. MMCTP uses an MLP to model each feature's overall trend, uses convolutional kernels of sizes 3, 5, and 7 in an MSCNN to capture local interactions at multiple resolutions, and then uses the local representation as queries and the global representation as keys and values in a multi-head cross-attention layer. A theoretical analysis decomposes the generalization error into variance, bias, threshold, and noise, and argues that omitting the global branch raises variance while omitting the local branch raises bias. Across input lengths from 24 to 192 steps and prediction horizons from 3 to 48 steps, and across 5s, 10s, and 15s sampling intervals, the paper reports that MMCTP attains the best MSE and MAE among the compared models, with the largest measured gains on the 15s interval.

Load-bearing premise

The load-bearing premise is that the three-seed averages reported for each model are stable enough that the 2–8% differences, especially the headline 5.04% MSE gap, reflect a real ordering rather than run-to-run training noise.

Editorial extensions

If this is right

  • If the claim holds, MMCTP is the best performing multi-step trajectory predictor among the compared models on GeoLife, across sampling intervals, input lengths, and prediction horizons.
  • The parallel global/local design predicts all future steps at once, so inference time stays nearly constant as the prediction horizon grows, unlike autoregressive baselines.
  • The $O(q)$ complexity means the fusion idea can be applied to long input sequences without the quadratic cost of standard attention.
  • Ablation results imply each component earns its place: removing the local branch, global branch, or fusion raises both MSE and MAE.
  • The largest relative gains occur on the 15s sampling interval, suggesting the method is especially useful when trajectory data is sparser.

Reading between the lines

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

  • An implication the authors do not develop: the headline 5.04% and 4.35% reductions are configuration-specific (15s sampling, 12-step prediction); on the 5s and 10s datasets the gains over ModernTCN are smaller, so the broad best-model claim rests mainly on the sparse-data setting.
  • A testable extension would replace zero-padded future positions in the local branch with iteratively generated predictions, since zero padding may damp the very local dynamics the MSCNN is meant to capture.
  • The bias-variance argument motivates keeping both branches but does not quantify the expected gain; the paper's case ultimately rests on the empirical ordering, so a multi-seed statistical comparison would strengthen it.
  • Because the model is $O(q)$ and predicts all steps in parallel, it is a natural candidate to plug into edge service migration pipelines, where latency budgets matter.
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

4 major / 7 minor

Summary. This paper proposes MMCTP, a neural architecture for multi-step GPS trajectory prediction on the GeoLife dataset. The model uses an MLP branch to extract global temporal patterns, a multi-scale CNN branch with convolutional kernel sizes 3, 5, and 7 to extract local temporal patterns over the most recent R time steps and zero-padded future positions, and a cross-attention module that uses the local representation as queries and the global representation as keys and values. Training uses RevIN normalization, an embedding layer, Huber loss, Adam optimization, and early stopping. Experiments compare MMCTP with LSTM, Seq2Seq LSTM, TCN, Informer, PatchTST, TSMixer, and ModernTCN across three sampling intervals, four input lengths, and five prediction horizons, with additional ablation and inference-time measurements. The paper reports up to 5.04% MSE and 4.35% MAE improvement over ModernTCN, claims O(q) time complexity, and offers a bias-variance argument for combining global and local branches.

Significance. The architecture is clearly specified and the experimental coverage is broad: three sampling intervals, four input lengths, five prediction horizons, seven baselines, ablation studies, and inference-time measurements. The evaluation uses a proper held-out test split, so the results are genuine predictions rather than fitted values. If the reported gains are statistically stable, MMCTP would be a simple, efficient state-of-the-art for the GeoLife trajectory-prediction benchmark and a useful reference design for fusing global and local temporal information. The main weakness is that the empirical claims rest on three-seed means with no variance information, and several effect sizes are small; the local branch's zero-padding also deserves scrutiny. The theoretical section is too qualitative to count as a proof of architectural necessity.

major comments (4)
  1. [§V-B, Tables V-X] Section V-B states that all experiments are repeated three times with different random seeds and the mean is used, but no standard deviation, per-seed value, or significance test is reported anywhere. This is load-bearing because many of the claimed improvements are small. For example, on the 5s dataset, MMCTP versus ModernTCN is 3.26% in MSE at input length 48 (Table V) and 2.09% at prediction length 48 (Table VIII); on the 10s dataset, the margin at prediction length 6 is 2.37% (Table IX); on the input-length-24 row of Table V, the margin is about 3.0%. With only three seeds, a 2-3% gap could easily be within run-to-run noise. The ablation table shows even smaller differences: on the 5s dataset, removing MLP changes MSE from 0.0005526 to 0.0005534 (0.14%) and using fixed-size kernels changes it to 0.0005532 (0.11%). Without variance information, the paper's central claim that MMCTP consistently outperforms ModernTCN is not statistically established. The authors should report per-seed results or standard deviations and, ideally, a paired significance test or bootstrap interval over seeds.
  2. [Algorithm 1, lines 9-10; §IV-C] The local branch concatenates the observed R recent time steps with n zero-valued future positions and then applies convolutions over the resulting R+n sequence. This is not standard padding; the zeros are inserted as data values, so every convolution window that reaches into the prediction region uses a constant-zero pattern. Because the fused representation I is used as queries in the cross-attention module, the zero region can directly influence the predicted outputs. The paper provides no experiment or argument showing that this padding does not bias the learned local features or the ablation results; for example, one would want a comparison against learned placeholder embeddings or a local branch that does not operate on the padded future region. As written, the 'local temporal information' extracted by MSCNN is partly an artifact of the zero placeholder.
  3. [§IV-E, Eq. (22)] The theoretical analysis does not prove that removing either branch degrades performance. The bias-variance decomposition in Eq. (22) silently assumes that the label noise has zero conditional mean when the cross term 2E[(fbar(x)-y)(y-y_psi)] is dropped, but this assumption is not stated. More importantly, scenarios (1) and (2) assert that removing global information increases variance and removing local information increases bias, but no formal link between the MLP/MSCNN modules and these quantities is given. The conclusion that 'the global information extraction component cannot be omitted' would apply to almost any two-part model under the same informal reasoning. This section should be reframed as intuition or replaced with a concrete analysis of the actual architecture.
  4. [§V-A] The experimental setup is ambiguous about the actual temporal resolution. The text says 'trajectories are extracted using sampling intervals of 5s, 10s, and 15s' but then 'linear interpolation is applied to reconstruct data points at 10s, 20s, and 30s intervals, respectively.' If data points are reconstructed at 10s intervals, the '5s sampling interval dataset' is not actually at 5s resolution, and the physical duration of a '48-step input / 12-step prediction' differs by a factor of two from what the labels suggest. This needs to be clarified or corrected, because it affects the interpretation of every table and the abstract's claim about '12-step prediction.'
minor comments (7)
  1. [Abstract and §V-D] The headline 'reduces MSE by 5.04% and MAE by 4.35% in 12-step prediction' is specific to the 15s sampling interval dataset with 48-step input (Tables VII and X); please qualify this in the abstract to avoid overstating the result.
  2. [§V-D] The statement that baseline hyperparameters are 'consistent with those used in the original papers' is not reproducible as written; please provide a table of the exact configuration used for each baseline or a reference to a released implementation.
  3. [§V-A, §V-B] The paper does not report the number of test samples, the specific random seeds, or whether the same train/validation/test split is used for all models; adding these details would improve reproducibility and help interpret the three-seed means.
  4. [§V-D, Figs. 7-8] Inference-time comparisons are shown only in figures; please include the numeric values in a table so readers can verify the claim that MMCTP's inference time is similar to ModernTCN's.
  5. [Eqs. (2), (29)-(30)] The notation for the prediction-step index is inconsistent: Eq. (2) sums over j=m+1 to m+n, while Eqs. (29) and (30) sum over j=1 to n with indices a_{i,m+j}; unify the notation.
  6. [§II, §IV-D] The phrase 'historical textual data' for GPS coordinate sequences is misleading and should be replaced with something like 'discrete GPS coordinate sequences.' Also, Eq. (20) uses an intermediate output tilde-y that is not defined in the text.
  7. [Fig. 1, Fig. 2] Fig. 1 defines graphical elements, but the caption of Fig. 2 does not map those elements to the MLP, MSCNN, and CA modules; please clarify the legend and figure caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: MMCTP is trained on a split and evaluated on a held-out split against external baselines; the few self-citations are not load-bearing.

full rationale

The paper's central claim is an empirical one: MMCTP (MLP + MSCNN + CA) reduces MSE/MAE on GeoLife compared with LSTM, Seq2Seq LSTM, TCN, Informer, PatchTST, TSMixer, and ModernTCN. The results in Tables V-X are obtained by training on the first 70% of each user's sub-trajectories, validating on the next 10%, and testing on the last 20% (Section V-A). The model is therefore predicting unseen future positions, not re-fitting or re-naming the training data. Hyperparameters (Table III) are fixed constants; no parameter is fitted to the test-set errors and then reported as a prediction. The only self-citations are references [1], [2], and [10] by author Chen Hou, used for edge-computing motivation and general neural-network control context; none is invoked to justify the MMCTP architecture or the claimed accuracy, so they are not load-bearing. The theoretical analysis in Section IV-E is a heuristic bias-variance argument about why both global and local information matter; it is qualitative and contains an unproven assertion that omitting a branch increases variance or bias, but this is a rigor/correctness concern rather than a circular derivation, since it does not define the model's output in terms of its inputs or rename a fitted quantity as a prediction. The paper also acknowledges in Section I that sudden route deviations are inherently difficult to model and are beyond its scope. No equation or algorithm step reduces to its own inputs by construction: in particular, the zero-padded future positions in Algorithm 1 (lines 9-10) are input placeholders for the convolutional branch, while the loss is computed against the held-out ground-truth future positions, so the prediction target is not fed into the model as a label. The absence of standard deviations or significance tests affects statistical reliability of the headline margins, but it is not a circularity issue. I find no circular step and assign a score of 0.

Assumptions & free parameters 9 free parameters · 4 assumptions · 0 invented entities

All free parameters are hyperparameters chosen by hand; the model weights are learned from data so they are not listed. No new physical or conceptual entities are introduced beyond standard neural network components.

free parameters (9)
  • MLP hidden size B = 2048
    Hand-chosen; controls capacity of global temporal MLP.
  • MSCNN/CA hidden size D = 256
    Hand-chosen; dimension of local branch and attention.
  • Local window length R = 24
    Hand-chosen; number of recent steps used for local extraction.
  • Convolutional kernel sizes = 3, 5, 7
    Hand-chosen; multi-scale temporal receptive fields.
  • Number of MSCNN blocks M = 2
    Hand-chosen; depth of local branch.
  • Number of MLP blocks N = 1
    Hand-chosen; depth of global branch.
  • Number of attention heads h = 8
    Hand-chosen; multi-head cross-attention.
  • Huber loss threshold delta = 0.001
    Hand-chosen; switch point for loss function.
  • Initial learning rate = 2e-5
    Hand-chosen; Adam optimizer.
assumptions (4)
  • standard math Jensen's inequality and bias-variance decomposition apply to the Huber loss with the stated residual definitions.
    Used in Section IV-E to decompose generalization error; these are classical results.
  • domain assumption The GeoLife dataset, after the described preprocessing (sampling intervals, linear interpolation, filtering users with fewer than 25 days), is representative of user mobility for evaluating trajectory prediction.
    Section V-A describes the dataset construction; the validity of results depends on this being a fair benchmark.
  • ad hoc to paper An MLP layer with a global receptive field captures long-range temporal patterns, while convolutional kernels of sizes 3, 5, 7 capture local patterns at different time scales, and these two representations are complementary for trajectory prediction.
    Sections IV-B and IV-C assert these inductive biases without independent justification; the entire architecture rests on this assumption.
  • ad hoc to paper Zero-padding the future prediction positions in the local branch does not introduce spurious information that materially biases the learned features.
    Algorithm 1 lines 9-12 concatenate zeroed future steps to the recent window; the paper provides no analysis of the effect of this padding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of User Trajectory Prediction Unifying Global and Local Temporal Information." pith.science (2026). https://pith.science/paper/N2UN7NYI

@misc{pith2026250802161,
  author       = {Pith},
  title        = {Pith review of: User Trajectory Prediction Unifying Global and Local Temporal Information},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N2UN7NYI}},
  note         = {Machine review of arXiv:2508.02161}
}
read the original abstract

Trajectory prediction is essential for formulating proactive strategies that anticipate user mobility and support advance preparation. Therefore, how to reduce the forecasting error in user trajectory prediction within an acceptable inference time arises as an interesting issue. However, trajectory data contains both global and local temporal information, complicating the extraction of the complete temporal pattern. Moreover, user behavior occurs over different time scales, increasing the difficulty of capturing behavioral patterns. To address these challenges, a trajectory prediction model based on multilayer perceptron (MLP), multi-scale convolutional neural network (MSCNN), and cross-attention (CA) is proposed. Specifically, MLP is used to extract the global temporal information of each feature. In parallel, MSCNN is employed to extract the local temporal information by modeling interactions among features within a local temporal range. Convolutional kernels with different sizes are used in MSCNN to capture temporal information at multiple resolutions, enhancing the model's adaptability to different behavioral patterns. Finally, CA is applied to fuse the global and local temporal information. Experimental results show that our model reduces mean squared error (MSE) by 5.04% and mean absolute error (MAE) by 4.35% compared with ModernTCN in 12-step prediction, while maintaining similar inference time.

Figures

Figures reproduced from arXiv: 2508.02161 by the authors.

Figure 1
Figure 1. Definitions of the graphical elements used in the figures. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The network architecture diagram of MMCTP. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The architecture diagram of the MLP module for global information extraction. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: The architecture diagram of the MSCNN module for local information extraction. The convolution operates on local regions defined by different [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The experimental architecture. TABLE I COMPARISON OF TIME COMPLEXITY Model Algorithm Time Complexity LSTM O(q) Seq2Seq LSTM O(q) TCN O(q) Informer O(q log q) ModernTCN O(q) PatchTST O(q 2 ) TSMixer O(q) MMCTP O(q) F. Time Complexity Analysis The time complexity of MMCT…
Figure 6
Figure 6. Figure 6: The experimental organization. TABLE III MODEL PARAMETER SETTINGS Parameter Value Optimizer Adam Loss Function Huber Loss δ 0.001 Dropout 0.05 Batch Size 32 Epoch 50 Patience Threshold 5 Initial Learning Rate 2e-5 N 1 MLP Activation Function ReLU B 2048 M 2 MSCNN Activ…
Figure 7
Figure 7. Figure 7: The impact of input length on inference time for various models on datasets with different sampling intervals. (a) 5s sampling interval dataset. (b) [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: The impact of prediction length on inference time for various models on datasets with different sampling intervals. (a) 5s sampling interval dataset. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: The visualization display of prediction results. (a) Predicted longitude. (b) Predicted latitude. (c) Predicted altitude. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 45 canonical work pages

  1. [1]

    Optimal task-offloading control for edge com- puting system with tasks offloaded and computed in sequence,

    C. Hou and Q. Zhao, “Optimal task-offloading control for edge com- puting system with tasks offloaded and computed in sequence,” IEEE Trans. Autom. Sci. Eng. , vol. 20, no. 2, pp. 1378–1392, 2022

    work page 2022

  2. [2]

    Optimal control of wireless powered edge computing system for balance between computation rate and energy harvested,

    C. Hou and Q. Zhao, “Optimal control of wireless powered edge computing system for balance between computation rate and energy harvested,” IEEE Trans. Autom. Sci. Eng. , vol. 20, no. 2, pp. 1108– 1124, 2022

    work page 2022

  3. [3]

    A trajectory prediction-based and dependency-aware container migration for mobile edge computing,

    W. Zhang, J. Luo, L. Chen, and J. Liu, “A trajectory prediction-based and dependency-aware container migration for mobile edge computing,” IEEE Trans. Serv. Comput. , vol. 16, no. 5, pp. 3168–3181, 2023

    work page 2023

  4. [4]

    Reinforced-LSTM trajectory prediction-driven dynamic service migration: A case study,

    Z. Zhao et al, “Reinforced-LSTM trajectory prediction-driven dynamic service migration: A case study,” IEEE Trans. Netw. Sci. Eng. , vol. 9, no. 4, pp. 2786–2802, 2022

    work page 2022

  5. [5]

    An attention-based spatiotem- poral LSTM network for next POI recommendation,

    L. Huang, Y . Ma, S. Wang, and Y . Liu, “An attention-based spatiotem- poral LSTM network for next POI recommendation,” IEEE Trans. Serv. Comput., vol. 14, no. 6, pp. 1585–1597, 2019

    work page 2019

  6. [6]

    SEABIG: A deep learning-based method for location prediction in pedestrian semantic trajectories,

    W. Zhang, L. Sun, X. Wang, Z. Huang, and B. Li, “SEABIG: A deep learning-based method for location prediction in pedestrian semantic trajectories,” IEEE Access, vol. 7, pp. 109 054–109 062, 2019

    work page 2019

  7. [7]

    Improving next location recommendation services with spatial-temporal multi-group contrastive learning,

    Z. Jia, Y . Fan, J. Zhang, C. Wei, R. Yan, and X. Wu, “Improving next location recommendation services with spatial-temporal multi-group contrastive learning,” IEEE Trans. Serv. Comput. , vol. 16, no. 5, pp. 3467–3478, 2023

    work page 2023

  8. [8]

    Mining interesting locations and travel sequences from GPS trajectories,

    Y . Zheng, L. Zhang, X. Xie, and W.-Y . Ma, “Mining interesting locations and travel sequences from GPS trajectories,” in Proc. 18th Int. Conf. World Wide Web, 2009, pp. 791–800

    work page 2009

Show all 52 references
  1. [9]

    Trajectory prediction of cyclists using a physical model and an artificial neural network,

    S. Zernetsch, S. Kohnen, M. Goldhammer, K. Doll, and B. Sick, “Trajectory prediction of cyclists using a physical model and an artificial neural network,” in Proc. 2016 IEEE Intell. Vehicles Symp. , 2016, pp. 833–838

    2016

  2. [10]

    Activation control of multiple piecewise linear neural net- works,

    C. Hou, “Activation control of multiple piecewise linear neural net- works,” IEEE Trans. Autom. Sci. Eng. , vol. 22, pp. 4895–4907, 2024

    2024

  3. [11]

    Social-aware trajec- tory prediction using goal-directed attention networks with egocentric vision,

    L. Astuti, C.-H. Chiu, Y .-C. Lin, and M.-C. Lin, “Social-aware trajec- tory prediction using goal-directed attention networks with egocentric vision,” PeerJ Comput. Sci., vol. 11, 2025, Art. no. e2842

    2025

  4. [12]

    STF-RNN: Space time features-based recurrent neural network for predicting people next location,

    A. Al-Molegi, M. Jabreel, and B. Ghaleb, “STF-RNN: Space time features-based recurrent neural network for predicting people next location,” in Proc. 2016 IEEE Symp. Ser. Comput. Intell. , 2016, pp. 1–7

    2016

  5. [13]

    Deep mobile path prediction with shift-and-join and carry-ahead,

    H. Yang, S. M. Raza, M. Kim, and H. Choo, “Deep mobile path prediction with shift-and-join and carry-ahead,” IEEE Trans. on Cogn. Commun. Netw., vol. 9, no. 3, pp. 811–825, 2023. 15

    2023

  6. [14]

    Base station handover based on user trajectory prediction in 5G networks,

    Y . Ma, X. Chen, and L. Zhang, “Base station handover based on user trajectory prediction in 5G networks,” in Proc. 2021 IEEE Int. Conf. Parallel Distrib. Process. Appl., Big Data Cloud Comput., Sustain. Comput. Commun., Social Comput. Netw. , 2021, pp. 1476–1482

    2021

  7. [15]

    Multiscale global and local self- attention-based network for remaining useful life prediction,

    Z. Zhang, W. Song, Q. Li, and H. Gao, “Multiscale global and local self- attention-based network for remaining useful life prediction,” Meas. Sci. Technol., vol. 34, no. 12, 2023, Art. no. 125154

    2023

  8. [16]

    Human motion trajectory prediction: A survey,

    A. Rudenko, L. Palmieri, M. Herman, K. M. Kitani, D. M. Gavrila, and K. O. Arras, “Human motion trajectory prediction: A survey,” Int. J. Robot. Res., vol. 39, no. 8, pp. 895–935, 2020

    2020

  9. [17]

    A two-layer switching based trajectory prediction method,

    S. Reisinger, D. Adelberger, and L. del Re, “A two-layer switching based trajectory prediction method,” Eur. J. Control , vol. 62, pp. 143–150, 2021

    2021

  10. [18]

    Map-based long term motion prediction for vehicles in traffic environments,

    D. Petrich, T. Dang, D. Kasper, G. Breuel, and C. Stiller, “Map-based long term motion prediction for vehicles in traffic environments,” in Proc. 16th Int. IEEE Conf. on Intell. Transp. Syst. , 2013, pp. 2166– 2172

    2013

  11. [19]

    Grip++: Enhanced graph-based interaction-aware trajectory prediction for autonomous driving,

    X. Li, X. Ying, and M. C. Chuah, “Grip++: Enhanced graph-based interaction-aware trajectory prediction for autonomous driving,” 2019, arXiv:1907.07792

    2019 arXiv

  12. [20]

    PoPPL: Pedestrian trajectory prediction by LSTM with automatic route class clustering,

    H. Xue, D. Q. Huynh, and M. Reynolds, “PoPPL: Pedestrian trajectory prediction by LSTM with automatic route class clustering,” IEEE Trans. Neural Netw. Learn. Syst. , vol. 32, no. 1, pp. 77–90, 2020

    2020

  13. [21]

    SocialCircle+: Learning the angle-based conditioned interaction representation for pedestrian trajectory prediction,

    C. Wong, B. Xia, Z. Zou, and X. You, “SocialCircle+: Learning the angle-based conditioned interaction representation for pedestrian trajectory prediction,” 2024, arXiv:2409.14984

    2024 arXiv

  14. [22]

    Hidden Markov model based user mobility analysis in LTE network,

    Q. Lv, Z. Mei, Y . Qiao, Y . Zhong, and Z. Lei, “Hidden Markov model based user mobility analysis in LTE network,” in Proc. 2014 Int. Symp. Wireless Pers. Multimedia Commun., 2014, pp. 379–384

    2014

  15. [23]

    Predicting personal transitional location based on modified-SVM,

    C. Xu and C. Xu, “Predicting personal transitional location based on modified-SVM,” in Proc. 2017 Int. Conf. Comput. Sci. Comput. Intell. , 2017, pp. 340–344

    2017

  16. [24]

    Multi-channel convolutional neural networks for han- dling multi-dimensional semantic trajectories and predicting future se- mantic locations,

    A. Karatzoglou, “Multi-channel convolutional neural networks for han- dling multi-dimensional semantic trajectories and predicting future se- mantic locations,” in Proc. Int. Workshop Multiple-Aspect Anal. Seman- tic Trajectories, 2019, pp. 117–132

    2019

  17. [25]

    AI-empowered VNF migration as a cost-loss-effective solution for network resilience,

    A. L. Ibrahimpa ˇsi´c, B. Han, and H. D. Schotten, “AI-empowered VNF migration as a cost-loss-effective solution for network resilience,” in Proc. 2021 IEEE Wireless Commun. Netw. Conf. Workshops , 2021, pp. 1–6

    2021

  18. [26]

    Mobile user trajectory prediction based on machine learning,

    Y . Liu, H. Yang, and R. Huang, “Mobile user trajectory prediction based on machine learning,” in Proc. 2022 IEEE 95th Veh. Technol. Conf. , 2022, pp. 1–5

    2022

  19. [27]

    ST-MoE-BERT: A spatial-temporal mixture-of-experts framework for long-term cross-city mobility predic- tion,

    H. He, H. Luo, and Q. R. Wang, “ST-MoE-BERT: A spatial-temporal mixture-of-experts framework for long-term cross-city mobility predic- tion,” in Proc. 2nd ACM SIGSPATIAL Int. Workshop Human Mobility Prediction Challenge, 2024, pp. 10–15

    2024

  20. [28]

    Exploring trajectory prediction through machine learning methods,

    C. Wang, L. Ma, R. Li, T. S. Durrani, and H. Zhang, “Exploring trajectory prediction through machine learning methods,” IEEE Access, vol. 7, pp. 101 441–101 452, 2019

    2019

  21. [29]

    A new framework for multi-hop ABS-assisted 5G-networks with users’ mobility prediction,

    E. Chaalal, S.-M. Senouci, and L. Reynaud, “A new framework for multi-hop ABS-assisted 5G-networks with users’ mobility prediction,” IEEE Trans. Veh. Technol., vol. 71, no. 4, pp. 4412–4427, 2022

    2022

  22. [30]

    Efficient and effective time-series forecasting with spiking neural networks,

    C. Lv, Y . Wang, D. Han, X. Zheng, X. Huang, and D. Li, “Efficient and effective time-series forecasting with spiking neural networks,” 2024, arXiv:2402.01533

    2024 arXiv

  23. [31]

    Attention-based spatiotemporal graph fusion convolution networks for water quality prediction,

    J. Qiao, Y . Lin, J. Bi, H. Yuan, G. Wang, and M. Zhou, “Attention-based spatiotemporal graph fusion convolution networks for water quality prediction,” IEEE Trans. Autom. Sci. Eng. , vol. 22, pp. 1–10, 2024

    2024

  24. [32]

    BERT: Pre-training of deep bidirectional Transformers for language understanding,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional Transformers for language understanding,” in Proc. 2019 Conf. North American Chapter Assoc. Comput. Linguistics: Human Lang. Technologies, 2019, pp. 4171–4186

    2019

  25. [33]

    Informer: Beyond efficient Transformer for long se- quence time-series forecasting,

    H. Zhou et al, “Informer: Beyond efficient Transformer for long se- quence time-series forecasting,” in Proc. 35th AAAI Conf. Artif. Intell. , vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  26. [34]

    A time series is worth 64 words: Long-term forecasting with Transformers,

    Y . Nie, N. H. Nguyen, P. Sinthong, and J. Kalagnanam, “A time series is worth 64 words: Long-term forecasting with Transformers,” 2022, arXiv:2211.14730

    2022 arXiv

  27. [35]

    Are Transformers effective for time series forecasting?

    A. Zeng, M. Chen, L. Zhang, and Q. Xu, “Are Transformers effective for time series forecasting?” in Proc. 37th AAAI Conf. Artif. Intell. , vol. 37, no. 9, 2023, pp. 11 121–11 128

    2023

  28. [36]

    TSmixer: An all-MLP architecture for time series forecasting,

    S.-A. Chen, C.-L. Li, N. Yoder, S. O. Arik, and T. Pfister, “TSmixer: An all-MLP architecture for time series forecasting,” 2023, arXiv:2303.06053

    2023 arXiv

  29. [37]

    MLog: Mogrifier LSTM-based log anomaly detection approach using semantic representation,

    Y . Fu, K. Liang, and J. Xu, “MLog: Mogrifier LSTM-based log anomaly detection approach using semantic representation,” IEEE Trans. Serv. Comput., vol. 16, no. 5, pp. 3537–3549, 2023

    2023

  30. [38]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” 2018, arXiv:1803.01271

    2018 arXiv

  31. [39]

    SCINet: Time series modeling and forecasting with sample convolution and interaction,

    M. Liu et al, “SCINet: Time series modeling and forecasting with sample convolution and interaction,” inProc. 36th Int. Conf. Neural Inf. Process. Syst., 2022, pp. 5816–5828

    2022

  32. [40]

    ModernTCN: A modern pure convolution structure for general time series analysis,

    D. Luo and X. Wang, “ModernTCN: A modern pure convolution structure for general time series analysis,” in Proc. 12th Int. Conf. Learn. Representations, 2024, pp. 1–43

    2024

  33. [41]

    Online-offline interactive urban crowd flow prediction toward IoT-based smart city,

    Y . Zeng, S. Zhou, and K. Xiang, “Online-offline interactive urban crowd flow prediction toward IoT-based smart city,”IEEE Trans. Serv. Comput., vol. 15, no. 6, pp. 3417–3428, 2021

    2021

  34. [42]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” 2014, arXiv:1412.3555

    2014 arXiv

  35. [43]

    Physics-informed neural network for rapid prediction of the temperature fields in fuel- heat pipe assemblies,

    D. Jiang, Z. Wang, L. Yuan, J. Gou, and J. Shan, “Physics-informed neural network for rapid prediction of the temperature fields in fuel- heat pipe assemblies,” Energy, vol. 332, 2025, Art. no. 137268

    2025

  36. [44]

    An interpretable hybrid TCN-BiLSTM model for reference evapotranspiration prediction,

    Z. Gao et al, “An interpretable hybrid TCN-BiLSTM model for reference evapotranspiration prediction,” Water Resour. Manage., pp. 1–23, 2025

    2025

  37. [45]

    Deep-learning model for pre- dicting physical fitness in possible sarcopenia: analysis of the korean physical fitness award from 2010 to 2023,

    J.-H. Bae, J.-W. Seo, and D. Y . Kim, “Deep-learning model for pre- dicting physical fitness in possible sarcopenia: analysis of the korean physical fitness award from 2010 to 2023,” Frontiers Public Health , vol. 11, 2023, Art. no. 1241388

    2010

  38. [46]

    Reversible instance normalization for accurate time-series forecasting against dis- tribution shift,

    T. Kim, J. Kim, Y . Tae, C. Park, J.-H. Choi, and J. Choo, “Reversible instance normalization for accurate time-series forecasting against dis- tribution shift,” in Proc. 10th Int. Conf. Learn. Representations , 2022, pp. 1–25

    2022

  39. [47]

    A Transformer-based framework for multivariate time series representation learning,

    G. Zerveas, S. Jayaraman, D. Patel, A. Bhamidipaty, and C. Eickhoff, “A Transformer-based framework for multivariate time series representation learning,” in Proc. 27th ACM SIGKDD Conf. Knowl. Discovery Data Mining, 2021, pp. 2114–2124

    2021

  40. [48]

    Attention is all you need,

    A. Vaswani et al, “Attention is all you need,” in Proc. 31st Int. Conf. Neural Inf. Process. Syst. , 2017, pp. 6000–6010

    2017

  41. [49]

    Root-mean-square error (RMSE) or mean absolute error (MAE): When to use them or not,

    T. O. Hodson, “Root-mean-square error (RMSE) or mean absolute error (MAE): When to use them or not,” Geoscientific Model Develop. , vol. 15, no. 14, pp. 5481–5487, 2022

    2022

  42. [50]

    An empirical investigation of bias and variance in time series forecasting: modeling considerations and error evaluation,

    V . L. Berardi and G. P. Zhang, “An empirical investigation of bias and variance in time series forecasting: modeling considerations and error evaluation,” IEEE Trans. Neural Netw. , vol. 14, no. 3, pp. 668–679, 2003

    2003

  43. [51]

    Refinements of Jensen’s inequality via majorization results with applications in the information theory,

    Y . Deng, H. Ullah, M. A. Khan, S. Iqbal, and S. Wu, “Refinements of Jensen’s inequality via majorization results with applications in the information theory,” J. Math., vol. 2021, 2021, Art. no. 1951799

    2021

  44. [52]

    Bias-variance tradeoff in machine learning: Theoretical formulation and implications to structural engineering ap- plications,

    X. Guan and H. Burton, “Bias-variance tradeoff in machine learning: Theoretical formulation and implications to structural engineering ap- plications,” Structures, vol. 46, pp. 17–30, 2022

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.