Pith. sign in

REVIEW 5 major objections 5 minor 68 references

Unified Spatial-Temporal Edge-Enhanced Graph Networks for Pedestrian Trajectory Prediction

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Unified spatial-temporal graph with edge-to-edge reasoning beats prior pedestrian trajectory predictors.

desk verdict A well-built architecture paper with a genuinely new combination; the SOTA claim needs multi-seed confirmation and a cleaner hyperparameter protocol. read the letter →

arxiv 2502.02504 v1 pith:I6U4HS3O submitted 2025-02-04 cs.CV cs.AI

Ruochen Li , Tanqiu Qiao , Stamos Katsigiannis , Zhanxing Zhu , Hubert P. H. Shum This is my paper · ORCID
classification cs.CVcs.AI
keywords pedestriantrajectorypredictionspatial-temporalgraphedgeHodgeLaplacianconvolutiontransformerencodersocialinteractionmodelingmotionforecasting
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

Pedestrian trajectory prediction usually splits the problem into spatial interactions among people at one instant and temporal dependencies of each person across time. The paper argues that this split misses high-order cross-time interactions, where one pedestrian's behavior at one moment influences another person's path at a later moment, and that these interactions become diluted by multi-step message passing. UniEdge instead builds one unified spatial-temporal graph by cutting observed trajectories into overlapping patches, so any two moments that can interact are connected directly as first-order relationships. On top of this graph it runs a dual-graph convolution that models both person-to-person interactions and edge-to-edge influence propagation, then decodes future positions with a transformer encoder. The authors report that this method outperforms prior state-of-the-art on the ETH, UCY, and SDD benchmarks.

What carries the argument

The central object is the unified ST graph with patches, combined with the edge graph built by the first-order boundary operator. The boundary operator turns each edge of the unified graph into a node of a new edge graph, connecting two edge-nodes when the original edges share an endpoint; the first-order Hodge Laplacian $L_1 = B_1^\top B_1 + B_2^\top B_2$ then provides a spectral filter (approximated by Laguerre polynomials) for updating edge features. This edge graph is what carries implicit influence propagation between interaction patterns. The node graph and edge graph are fused in a GCN where normalized edge embeddings act as weights in node aggregation. The patch construction also reduces the number of transformer tokens from $T_{obs}$ to $K$, giving a quadratic reduction in attention cost.

What would settle it

Keep UniEdge fixed and replace the edge graph used in E2E-N2N-GCN with a randomly rewired graph that has the same number of nodes, edges, and feature dimensions, or with a fully connected edge graph. If ADE/FDE on ETH/UCY stay the same, then the boundary-operator topology is not the source of the improvement; if they degrade, the topology carries the effect.

Watch

Extended reading notes

Core claim

UniEdge claims that pedestrian trajectory prediction improves when high-order cross-time interactions are flattened into first-order edges of a unified spatial-temporal graph, and when implicit interactions between interaction patterns are modeled through an edge graph. The paper constructs patch-based unified ST graphs so that any pair of pedestrian-time nodes appearing in the same patch are directly connected, avoiding multi-step aggregation and lowering effective resistance. The E2E-N2N-GCN then applies a first-order boundary operator to obtain an edge graph whose nodes are the original graph's edges, and convolves edge features with a Hodge-Laplacian Laguerre filter, allowing interaction patterns to influence each other. A transformer encoder with learnable future placeholders provides global temporal modeling. On the ETH, UCY, and Stanford Drone datasets under the best-of-20 protocol, UniEdge reports average ADE/FDE of 0.18/0.25 on ETH/UCY and 7.51/10.89 on SDD, surpassing the listed prior methods.

Load-bearing premise

The load-bearing premise is that the edge graph built from shared endpoints is what captures implicit influence propagation between interaction patterns, and the paper only supports this by removing the whole edge module rather than isolating the topology from added parameters.

Editorial extensions

If this is right

  • If the central claim is right, pedestrian trajectory models should stop separating spatial and temporal aggregation and instead connect cross-time interaction pairs directly, lowering effective resistance between distant nodes.
  • Edge features should be treated as first-class carriers of information, not just weights for node aggregation, because interaction patterns influence each other.
  • The transformer-encoder predictor with learnable future tokens shows that global temporal modeling can replace auto-regressive decoding without a full encoder-decoder stack, at lower FLOPs than comparable transformer methods.
  • The reported gains are largest in dense scenes like UNIV, suggesting that unified cross-time graphs matter most where many pedestrians interact closely.
  • Patch size and stride control the granularity of cross-time interaction, with patch size 3 and stride 1 reported as optimal.

Reading between the lines

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

  • The paper's effective-resistance argument could be tested directly by measuring $R_{ij}$ on real trajectory graphs before and after patching; the claim predicts a systematic drop for distant pedestrian-time node pairs.
  • Because the edge graph is built topologically from shared endpoints, a natural extension is to weight the edge graph by interaction strength instead of binary adjacency, which might improve sparse scenes like ETH.
  • The best-of-20 evaluation protocol rewards coverage, so the unified ST graph may be helping primarily by sharpening the predicted distribution rather than by top-1 accuracy; reporting minADE versus meanADE would clarify this.
  • If the line-graph topology is what matters, replacing HLLConv with any parameter-matched edge convolution that respects the same adjacency should not change results; the paper only ablates the whole edge module, not the topology.
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

5 major / 5 minor

Summary. The paper proposes UniEdge, a pedestrian trajectory prediction architecture with three main components: a patch-based unified spatial-temporal graph that turns high-order cross-time interactions into first-order relations, a dual-graph E2E-N2N-GCN module that combines node-level social interaction modeling with edge-level influence propagation via Hodge-Laplacian Laguerre convolution, and a transformer encoder-based predictor with learnable future placeholders. The authors report state-of-the-art average ADE/FDE of 0.18/0.25 on ETH/UCY and 7.51/10.89 on SDD under the best-of-20 protocol, together with component ablations, qualitative visualizations, and complexity comparisons.

Significance. If the reported gains are robust, the unified ST graph construction is a useful practical idea: it reduces the number of temporal tokens for the transformer predictor and it offers a concrete way to reduce effective resistance in message passing. The edge-level propagation mechanism is also a genuinely different design choice from the N2N and E2N paradigms common in trajectory prediction. The paper follows the standard bench protocol, compares against many recent baselines, and releases source code, which are strengths. However, the central SOTA claim is currently supported by single-run numbers, by hyperparameters (patch length and stride) selected on test data, and by an underspecified graph construction, so the evidence is not yet conclusive. The architecture is plausible, and the issues are fixable, but the experimental and specification gaps are load-bearing for the claimed contribution.

major comments (5)
  1. [§III-B, §III-C, Eqs. (2)-(8)] The construction of the node adjacency matrix A^k_node is never defined. Section III-B only states that A^k_node ∈ R^{NL×NL} 'encapsulates the node connections,' but the paper does not specify which pairs of nodes are connected within a patch, whether edges are spatial, temporal, or both, whether a distance threshold or k-nearest-neighbor rule is used, or how self-loops are handled. This is load-bearing because the edge graph G^k_edge is derived from this adjacency through the boundary operator, and Eq. (8) indexes Φ(H^k_{edge,ij}) even though the mapping from edge-graph node embeddings back to original edge features (i,j) is not specified. Without these definitions, the method cannot be reproduced or independently evaluated even with the released code.
  2. [§IV-E4, Fig. 9] Patch length L and stride S are selected by evaluating the final ADE/FDE on the ETH/UCY test subsets, and no validation split is described in Section IV-E4 or in the implementation details. Since L and S control the graph topology and the number of transformer tokens, tuning them on the test sets means the reported average gains are partially a result of test-set peeking. Please report L and S chosen on a validation split, or run a nested leave-one-out procedure, and show the corresponding final test results.
  3. [§IV-C, Tables I-II] All quantitative results are from a single run, with no standard deviations, seed counts, or significance tests. This matters because the aggregate margins are small on some subsets: on ETH, UniEdge's ADE (0.36) is worse than MSRL (0.28) and DDL (0.26), and the overall average improvement of 0.01-0.02 ADE over the next-best methods is comparable to typical seed-to-seed variation in this protocol. Please report results over at least 3-5 seeds, with means and standard deviations, and ideally paired statistical tests for the key comparisons.
  4. [§III-C, Table III] The ablation for the E2E component does not isolate the topological contribution of the edge graph. Variant (3) removes HLLConv entirely, which simultaneously removes parameters and feature channels, so the observed drop does not prove that the line-graph/Hodge-Laplacian structure is what helps. A controlled comparison with a same-capacity alternative, such as an MLP or GCN applied directly to edge features without the line-graph topology, is needed to support the claimed importance of E2E influence propagation.
  5. [§III-A] The preprocessing description is potentially problematic: the paper says it 'subtract[s] each historical v_t, t∈[−T_obs,0], by the corresponding endpoint v_Tpred,' where v_Tpred would be a future endpoint velocity that is not available at inference time. If v_Tpred is a typo and the intended subtraction uses the last observed velocity, please correct the notation; otherwise, this appears to leak ground-truth future information into the input features and would invalidate the evaluation protocol.
minor comments (5)
  1. [Abstract] The source code link says 'UniEge' instead of 'UniEdge'; please fix the typo.
  2. [§III-C, Algorithm 1] Please report the number J of Laguerre polynomial terms used in the experiments, as this hyperparameter is not listed in Section III-E.
  3. [§IV-E1, Table III] Variant (3) is described as 'without learning edge graphs through HLLConv'; please state explicitly what the model uses instead of HLLConv, so the ablation is unambiguous.
  4. [§IV-E2, Table V] The Euclidean-distance edge feature is called 'ours' in the text but is not formally defined in Section III; please specify exactly how E^k is formed from node coordinates and A^k_node, and report the values of σ and ε for the other kernels.
  5. [§IV-A] The leave-one-out protocol is standard, but the paper should state explicitly whether the reported results for the SDD datasets also use the best-of-20 sampling convention or a different sampling rule, since SDD numbers in the literature are not always directly comparable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: UniEdge's architecture is empirically evaluated, and no claimed result reduces to a fitted parameter or self-citation chain.

full rationale

This is an empirical architecture paper rather than a derivation, and the comparison against prior methods is based on standard train/test evaluation. The core components—the patch-based unified ST graph, the E2E-N2N-GCN with HLLConv, and the transformer encoder predictor—are defined constructively and tested through ablations on the benchmark datasets; no central claim is obtained by construction from a fitted constant or by renaming an input quantity as a prediction. Self-citations such as [8] and [35] appear only as related-work context and are not load-bearing for the main result. The HLLConv and first-order boundary operator are imported from external prior work ([25], [26], [28]) rather than from the authors' own unverified claims, so there is no self-citation chain forcing the conclusion. The hyperparameter study for patch size and stride (Section IV-E4, Fig. 9) is performed on the same test sets used for the final reported numbers, which is a benchmark-selection concern that may inflate the reported margins, but it is not circularity: the reported ADE/FDE values are measured outputs, not quantities that are definitionally equal to the selected hyperparameters. Overall, the paper's claims rest on empirical comparisons and code release, not on a derivation that presupposes its own conclusion.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The central claim rests on multiple hyperparameters selected on the test benchmarks, plus assumptions about the meaningfulness of topological graph constructions. The paper introduces no new physical or mathematical entities.

free parameters (5)
  • Patch length L = 3
    Selected via ablation on ETH/UCY test sets (Fig. 9, left). The paper reports that L=3 gives optimal ADE/FDE.
  • Stride S = 1
    Selected via ablation on ETH/UCY test sets (Fig. 9, right). S=1 gives the best performance due to increased number of patches.
  • Edge feature type = Euclidean distance kernel
    Chosen empirically from Table V, which compares reciprocal distance, Gaussian kernel, and Euclidean distance. The Euclidean kernel yields the best results.
  • Number of Laguerre polynomial terms J
    Not explicitly specified in the paper; a hyperparameter of HLLConv that the authors do not report or justify.
  • Transformer hyperparameters (hidden dim, heads, layers) = hidden 256, 4 heads
    Reported in Implementation Details but no sensitivity analysis is provided; these are standard architecture choices.
assumptions (3)
  • domain assumption Effective resistance (Eq. 1) and the under-reaching phenomenon apply to the unified ST graph as described.
    The paper motivates the unified graph by claiming it reduces effective resistance, but this is only illustrated on a toy graph (Fig. 4) and not measured on the actual model.
  • domain assumption The first-order boundary operator and Hodge Laplacian (Eq. 5) can be meaningfully applied to a graph derived from pedestrian trajectories.
    The construction is borrowed from topological data analysis, and the paper does not provide evidence that the line-graph topology captures real interaction-influence patterns.
  • domain assumption The evaluation protocol (best-of-20 sampling, leave-one-out splits for ETH/UCY) is a valid measure of prediction quality.
    The paper follows standard practice in the field, but best-of-20 metrics can be noisy and are not invariant to sampling, so reliance on this metric is an assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unified Spatial-Temporal Edge-Enhanced Graph Networks for Pedestrian Trajectory Prediction." pith.science (2026). https://pith.science/paper/I6U4HS3O

@misc{pith2026250202504,
  author       = {Pith},
  title        = {Pith review of: Unified Spatial-Temporal Edge-Enhanced Graph Networks for Pedestrian Trajectory Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I6U4HS3O}},
  note         = {Machine review of arXiv:2502.02504}
}
read the original abstract

Pedestrian trajectory prediction aims to forecast future movements based on historical paths. Spatial-temporal (ST) methods often separately model spatial interactions among pedestrians and temporal dependencies of individuals. They overlook the direct impacts of interactions among different pedestrians across various time steps (i.e., high-order cross-time interactions). This limits their ability to capture ST inter-dependencies and hinders prediction performance. To address these limitations, we propose UniEdge with three major designs. Firstly, we introduce a unified ST graph data structure that simplifies high-order cross-time interactions into first-order relationships, enabling the learning of ST inter-dependencies in a single step. This avoids the information loss caused by multi-step aggregation. Secondly, traditional GNNs focus on aggregating pedestrian node features, neglecting the propagation of implicit interaction patterns encoded in edge features. We propose the Edge-to-Edge-Node-to-Node Graph Convolution (E2E-N2N-GCN), a novel dual-graph network that jointly models explicit N2N social interactions among pedestrians and implicit E2E influence propagation across these interaction patterns. Finally, to overcome the limited receptive fields and challenges in capturing long-range dependencies of auto-regressive architectures, we introduce a transformer encoder-based predictor that enables global modeling of temporal correlation. UniEdge outperforms state-of-the-arts on multiple datasets, including ETH, UCY, and SDD.

Figures

Figures reproduced from arXiv: 2502.02504 by the authors.

Figure 1
Figure 1. Motivation Illustration. (a) Real-world pedestrian trajectories over multiple time frames. (b) Existing ST approaches separately model the spatial interactions among pedestrians and temporal dependencies of individuals. (c) Our unified ST graph integrates ST inter-dependencies and simplifies high-order cross-time interactions into first-order relationships. Predicting pedestrian trajectory is inherently challenging,… view at source ↗
Figure 2
Figure 2. Illustration of graph learning procedures. (a) Node-to-Node (N2N), (b) Edge-to-Node (E2N), and (c) Our novel dual-graph introduces the com￾bination of N2N and Edge-to-Edge (E2E) paradigm. world scenarios [21], [22], particularly in situations requiring immediate response to dynamic changes. Another challenge lies in modeling the implicit influence propagation through edges in pedestrian social interactions. While Gr… view at source ↗
Figure 3
Figure 3. Overview of the proposed UniEdge. (a) Construction of patch-based unified ST graphs that simplify cross-time interactions into first-order relationships, (b) Edge-to-Edge-Node-to-Node Graph Convolution (E2E-N2N-GCN) that jointly processes N2N interactions and E2E influence propagation, and (c) Transformer Encoder-based trajectory predictor [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison of effective resistance (Rij ) between traditional ST approach (left, Rij = 1.50) and our unified ST graph (right, Rij = 0.27). Lower Rij indicates better message propagation efficiency. interactions, which require multi-step aggregation. Such multi￾step pro…
Figure 5
Figure 5. Figure 5: Illustration of edge graph construction from a unified ST graph using [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Illustration of the Transformer encoder-based predictor. [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Visualization of predicted trajectories on the ETH and UCY datasets. Historical trajectories are in blue, ground-truth trajectories are in red, and predicted trajectories are in yellow. Scenario (a) shows two pedestrians walking in parallel and meet; Scenario (b) illus…
Figure 8
Figure 8. Figure 8: Visualization of predicted distributions on the ETH and UCY datasets. Historical trajectories are in blue, ground-truth trajectories are in red, and predicted trajectories are in yellow. Scenario (a) and (b) show two pedestrians walking in parallel with convergence; (c…
Figure 10
Figure 10. Figure 10: Edge weight visualization of traditional two-stage ST approach [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 12
Figure 12. Figure 12: Sample scenario in ETH dataset. Historical trajectories are in blue, ground-truth trajectories are in red. shows how UniEdge constructs a unified ST graph between Ped.1 and Ped.2, even though their trajectories are relatively stable with minimal interaction, potential…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

68 extracted references · 60 canonical work pages

  1. [1]

    Intention-aware online pomdp planning for autonomous driving in a crowd,

    H. Bai, S. Cai, N. Ye, D. Hsu, and W. S. Lee, “Intention-aware online pomdp planning for autonomous driving in a crowd,” in IEEE Int. Conf. Robot. Autom. IEEE, 2015, pp. 454–460

    work page 2015

  2. [2]

    Multimodal pedestrian trajectory prediction using probabilistic proposal network,

    W. Chen, Z. Yang, L. Xue, J. Duan, H. Sun, and N. Zheng, “Multimodal pedestrian trajectory prediction using probabilistic proposal network,” IEEE Trans. Circuits Syst. Video Technol. , vol. 33, no. 6, pp. 2877– 2891, 2023

    work page 2023

  3. [3]

    Reciprocal twin networks for pedestrian motion learning and future path prediction,

    H. Sun, Z. Zhao, Z. Yin, and Z. He, “Reciprocal twin networks for pedestrian motion learning and future path prediction,” IEEE Trans. Circuits Syst. Video Technol., vol. 32, no. 3, pp. 1483–1497, 2021

    work page 2021

  4. [4]

    Prediction of pedestrian crossing behavior based on surveillance video,

    X. Zhou, H. Ren, T. Zhang, X. Mou, Y . He, and C.-Y . Chan, “Prediction of pedestrian crossing behavior based on surveillance video,” Sensors, 2022

    work page 2022

  5. [5]

    Trajectorycnn: a new spatio-temporal feature learning network for human motion prediction,

    X. Liu, J. Yin, J. Liu, P. Ding, J. Liu, and H. Liu, “Trajectorycnn: a new spatio-temporal feature learning network for human motion prediction,” IEEE Trans. Circuits Syst. Video Technol., vol. 31, no. 6, pp. 2133–2146, 2020

    work page 2020

  6. [6]

    Exploring spatio–temporal graph convolution for video- based human–object interaction recognition,

    N. Wang, G. Zhu, H. Li, M. Feng, X. Zhao, L. Ni, P. Shen, L. Mei, and L. Zhang, “Exploring spatio–temporal graph convolution for video- based human–object interaction recognition,” IEEE Trans. Circuits Syst. Video Technol., vol. 33, no. 10, pp. 5814–5827, 2023

    work page 2023

  7. [7]

    Sgcn: Sparse graph convolution network for pedestrian trajectory prediction,

    L. Shi, L. Wang, C. Long, S. Zhou, M. Zhou, Z. Niu, and G. Hua, “Sgcn: Sparse graph convolution network for pedestrian trajectory prediction,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2021, pp. 8994– 9003

    work page 2021

  8. [8]

    Multiclass-sgcn: Sparse graph- based trajectory prediction with agent class embedding,

    R. Li, S. Katsigiannis, and H. P. Shum, “Multiclass-sgcn: Sparse graph- based trajectory prediction with agent class embedding,” in IEEE Int. Conf. Image Process. IEEE, 2022, pp. 2346–2350

    work page 2022

Show all 68 references
  1. [9]

    Social-bigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks,

    V . Kosaraju, A. Sadeghian, R. Mart ´ın-Mart´ın, I. Reid, H. Rezatofighi, and S. Savarese, “Social-bigat: Multimodal trajectory forecasting using bicycle-gan and graph attention networks,” Proc. Adv. Neu. Inf. Process. Syst., vol. 32, 2019

    2019

  2. [10]

    Stgat: Modeling spatial-temporal interactions for human trajectory prediction,

    Y . Huang, H. Bi, Z. Li, T. Mao, and Z. Wang, “Stgat: Modeling spatial-temporal interactions for human trajectory prediction,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2019, pp. 6272–6281

    2019

  3. [11]

    Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction,

    A. Mohamed, K. Qian, M. Elhoseiny, and C. Claudel, “Social-stgcnn: A social spatio-temporal graph convolutional neural network for human trajectory prediction,” in Proc. IEEE Conf. Comput. Vis. Pattern Recog- nit., 2020, pp. 14 424–14 432

    2020

  4. [12]

    Learning pedestrian group repre- sentations for multi-modal trajectory prediction,

    I. Bae, J.-H. Park, and H.-G. Jeon, “Learning pedestrian group repre- sentations for multi-modal trajectory prediction,” in Proc. Eur. Conf. Comput. Vis. Springer, 2022, pp. 270–289

    2022

  5. [13]

    Eigentrajectory: Low-rank descriptors for multi-modal trajectory forecasting,

    I. Bae, J. Oh, and H.-G. Jeon, “Eigentrajectory: Low-rank descriptors for multi-modal trajectory forecasting,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023

    2023

  6. [14]

    A set of control points conditioned pedestrian trajectory prediction,

    I. Bae and H.-G. Jeon, “A set of control points conditioned pedestrian trajectory prediction,” in Proc. AAAI Conf. Artif. Intell. , vol. 37, no. 5, 2023, pp. 6155–6165

    2023

  7. [15]

    Graph Attention Networks,

    P. Veli ˇckovi´c, G. Cucurull, A. Casanova, A. Romero, P. Li`o, and Y . Ben- gio, “Graph Attention Networks,” Proc. Int. Conf. Learn. Represent. , 2018

    2018

  8. [16]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  9. [17]

    Semi-supervised classification with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proc. Int. Conf. Learn. Represent. , 2017

    2017

  10. [18]

    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,” arXiv preprint arXiv:1803.01271, 2018

    2018 arXiv

  11. [19]

    Nodemixup: Tackling under-reaching for graph neural networks,

    W. Lu, Z. Guan, W. Zhao, Y . Yang, and L. Jin, “Nodemixup: Tackling under-reaching for graph neural networks,” in Proc. AAAI Conf. Artif. Intell., vol. 38, no. 13, 2024, pp. 14 175–14 183

    2024

  12. [20]

    Understanding over- squashing in gnns through the lens of effective resistance,

    M. Black, Z. Wan, A. Nayyeri, and Y . Wang, “Understanding over- squashing in gnns through the lens of effective resistance,” in Proc. Int. Conf. Mach. Learn. PMLR, 2023, pp. 2528–2547

    2023

  13. [21]

    Fully- connected spatial-temporal graph for multivariate time-series data,

    Y . Wang, Y . Xu, J. Yang, M. Wu, X. Li, L. Xie, and Z. Chen, “Fully- connected spatial-temporal graph for multivariate time-series data,” in Proc. AAAI Conf. Artif. Intell., vol. 38, no. 14, 2024, pp. 15 715–15 724

    2024

  14. [22]

    FourierGNN: Rethinking multivariate time series forecasting from a pure graph perspective,

    K. Yi, Q. Zhang, W. Fan, H. He, L. Hu, P. Wang, N. An, L. Cao, and Z. Niu, “FourierGNN: Rethinking multivariate time series forecasting from a pure graph perspective,” in Proc. Adv. Neu. Inf. Process. Syst. , 2023

    2023

  15. [23]

    Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning,

    C. Xu, M. Li, Z. Ni, Y . Zhang, and S. Chen, “Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning,” Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , pp. 6488–6497, 2022

    2022

  16. [24]

    Heterogeneous edge-enhanced graph attention network for multi-agent trajectory prediction,

    X. Mo, Y . Xing, and C. Lv, “Heterogeneous edge-enhanced graph attention network for multi-agent trajectory prediction,” arXiv preprint arXiv:2106.07161, 2021

    2021 arXiv

  17. [25]

    Deciphering spatio-temporal graph forecasting: A causal lens and treatment,

    Y . Xia, Y . Liang, H. Wen, X. Liu, K. Wang, Z. Zhou, and R. Zimmer- mann, “Deciphering spatio-temporal graph forecasting: A causal lens and treatment,” in Proc. Adv. Neu. Inf. Process. Syst. , 2023

    2023

  18. [26]

    Heterogeneous graph convolutional neural network via hodge-laplacian for brain functional data,

    J. Huang, M. K. Chung, and A. Qiu, “Heterogeneous graph convolutional neural network via hodge-laplacian for brain functional data,” in Int. Conf. Inf. Process. Med. Imaging . Springer, 2023, pp. 278–290

    2023

  19. [27]

    Deep dual graph attention auto-encoder for community detection,

    X. Wu, W. Lu, Y . Quan, Q. Miao, and P. G. Sun, “Deep dual graph attention auto-encoder for community detection,” Expert Syst. Appl., vol. 238, p. 122182, 2024

    2024

  20. [28]

    First-order operators and boundary triples,

    O. Post, “First-order operators and boundary triples,” Russian Journal of Mathematical Physics , vol. 14, no. 4, pp. 482–492, 2007

    2007

  21. [29]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

    2017

  22. [30]

    You’ll never walk alone: Modeling social behavior for multi-target tracking,

    S. Pellegrini, A. Ess, K. Schindler, and L. Van Gool, “You’ll never walk alone: Modeling social behavior for multi-target tracking,” in Proc. IEEE/CVF Int. Conf. Comput. Vis. , 2009, pp. 261–268

    2009

  23. [31]

    Crowds by example,

    A. Lerner, Y . Chrysanthou, and D. Lischinski, “Crowds by example,” in Computer graphics forum , vol. 26, no. 3. Wiley Online Library, 2007, pp. 655–664

    2007

  24. [32]

    Learning social etiquette: Human trajectory understanding in crowded scenes,

    A. Robicquet, A. Sadeghian, A. Alahi, and S. Savarese, “Learning social etiquette: Human trajectory understanding in crowded scenes,” in Proc. Eur. Conf. Comput. Vis. Springer, 2016, pp. 549–565

    2016

  25. [33]

    Social lstm: Human trajectory prediction in crowded spaces,

    A. Alahi, K. Goel, V . Ramanathan, A. Robicquet, L. Fei-Fei, and S. Savarese, “Social lstm: Human trajectory prediction in crowded spaces,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2016, pp. 961–971

    2016

  26. [34]

    Social gan: Socially acceptable trajectories with generative adversarial networks,

    A. Gupta, J. Johnson, L. Fei-Fei, S. Savarese, and A. Alahi, “Social gan: Socially acceptable trajectories with generative adversarial networks,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 2255–2264

    2018

  27. [35]

    Geometric features informed multi-person human-object interaction recognition in videos,

    T. Qiao, Q. Men, F. W. B. Li, Y . Kubotani, S. Morishima, and H. P. H. Shum, “Geometric features informed multi-person human-object interaction recognition in videos,” in Proc. Eur. Conf. Comput. Vis. , 2022

    2022

  28. [36]

    Spatial temporal graph convolutional networks for skeleton-based action recognition,

    S. Yan, Y . Xiong, and D. Lin, “Spatial temporal graph convolutional networks for skeleton-based action recognition,” in Proceedings of the AAAI conference on artificial intelligence , vol. 32, no. 1, 2018

    2018

  29. [37]

    Skeleton-based human ac- tion recognition via large-kernel attention graph convolutional network,

    Y . Liu, H. Zhang, Y . Li, K. He, and D. Xu, “Skeleton-based human ac- tion recognition via large-kernel attention graph convolutional network,” IEEE Trans. Vis. Comput. Graph. , vol. 29, no. 5, pp. 2575–2585, 2023

    2023

  30. [38]

    Multiphysical graph neural network (mp-gnn) for covid-19 drug design,

    X.-S. Li, X. Liu, L. Lu, X.-S. Hua, Y . Chi, and K. Xia, “Multiphysical graph neural network (mp-gnn) for covid-19 drug design,” Briefings in bioinformatics, vol. 23, no. 4, p. bbac231, 2022

    2022

  31. [39]

    Neural graph collaborative filtering,

    X. Wang, X. He, M. Wang, F. Feng, and T.-S. Chua, “Neural graph collaborative filtering,” in Proc.Int. ACM SIGIR Conf. Res. Dev. Inf. Retrieval, 2019, pp. 165–174

    2019

  32. [40]

    Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,

    B. Yu, H. Yin, and Z. Zhu, “Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting,” in Int. Joint Conf. Artif. Intell. , 2018

    2018

  33. [41]

    Trajectory unified transformer for pedestrian trajectory prediction,

    L. Shi, L. Wang, S. Zhou, and G. Hua, “Trajectory unified transformer for pedestrian trajectory prediction,” in Proc. IEEE/CVF Int. Conf. Comput. Vis., 2023, pp. 9675–9684. 14

    2023

  34. [42]

    Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction,

    Y . Xu, A. Bazarjani, H.-g. Chi, C. Choi, and Y . Fu, “Uncovering the missing pattern: Unified framework towards trajectory imputation and prediction,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2023, pp. 9632–9643

    2023

  35. [43]

    Mfan: Mixing feature attention network for trajectory prediction,

    J. Li, L. Yang, Y . Chen, and Y . Jin, “Mfan: Mixing feature attention network for trajectory prediction,” Pattern Recognition , vol. 146, p. 109997, 2024

    2024

  36. [44]

    Socialvae: Human trajectory prediction using timewise latents,

    P. Xu, J.-B. Hayet, and I. Karamouzas, “Socialvae: Human trajectory prediction using timewise latents,” in Proc. Eur. Conf. Comput. Vis. , 2022, pp. 511–528

    2022

  37. [45]

    Aut- ofocusing for synthetic aperture imaging based on pedestrian trajectory prediction,

    Z. Pei, J. Zhang, W. Zhang, M. Wang, J. Wang, and Y .-H. Yang, “Aut- ofocusing for synthetic aperture imaging based on pedestrian trajectory prediction,” IEEE Trans. Circuits Syst. Video Technol. , vol. 34, no. 5, pp. 3551–3562, 2024

    2024

  38. [46]

    Context-aware human trajectories prediction via latent variational model,

    A. D ´ıaz Berenguer, M. Alioscha-Perez, M. C. Oveneke, and H. Sahli, “Context-aware human trajectories prediction via latent variational model,” IEEE Trans. Circuits Syst. Video Technol , vol. 31, no. 5, pp. 1876–1889, 2021

    2021

  39. [47]

    Mrgtraj: A novel non- autoregressive approach for human trajectory prediction,

    Y . Peng, G. Zhang, J. Shi, X. Li, and L. Zheng, “Mrgtraj: A novel non- autoregressive approach for human trajectory prediction,” IEEE Trans. Circuits Syst. Video Technol., vol. 34, no. 4, pp. 2318–2331, 2024

    2024

  40. [48]

    Pedestrian trajectory prediction using dynamics-based deep learning,

    H. Wang, W. Zhi, G. Batista, and R. Chandra, “Pedestrian trajectory prediction using dynamics-based deep learning,” in IEEE Int. Conf. Robot. Autom. IEEE, 2024, pp. 15 068–15 075

    2024

  41. [49]

    Minimizing effective resistance of a graph,

    A. Ghosh, S. Boyd, and A. Saberi, “Minimizing effective resistance of a graph,” SIAM review, vol. 50, no. 1, pp. 37–66, 2008

    2008

  42. [50]

    The moore–penrose inverse of the normalized graph lapla- cian,

    E. Bozzo, “The moore–penrose inverse of the normalized graph lapla- cian,” Linear Algebra Appl. , vol. 439, no. 10, pp. 3038–3043, 2013

    2013

  43. [51]

    Imagenet classification with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification with deep convolutional neural networks,” Adv. Neural Inf. Process. Syst., vol. 25, 2012

    2012

  44. [52]

    How attentive are graph attention networks?

    S. Brody, U. Alon, and E. Yahav, “How attentive are graph attention networks?” arXiv preprint arXiv:2105.14491 , 2021

    2021 arXiv

  45. [53]

    Multi- stream representation learning for pedestrian trajectory prediction,

    Y . Wu, L. Wang, S. Zhou, J. Duan, G. Hua, and W. Tang, “Multi- stream representation learning for pedestrian trajectory prediction,” in Proc. AAAI Conf. Artif. Intell. , vol. 37, no. 3, 2023, pp. 2875–2882

    2023

  46. [54]

    Autoregressive image gen- eration without vector quantization,

    T. Li, Y . Tian, H. Li, M. Deng, and K. He, “Autoregressive image gen- eration without vector quantization,” arXiv preprint arXiv:2406.11838 , 2024

    2024 arXiv

  47. [55]

    itrans- former: Inverted transformers are effective for time series forecasting,

    Y . Liu, T. Hu, H. Zhang, H. Wu, S. Wang, L. Ma, and M. Long, “itrans- former: Inverted transformers are effective for time series forecasting,” arXiv preprint arXiv:2310.06625 , 2023

    2023 arXiv

  48. [56]

    Social- implicit: Rethinking trajectory prediction evaluation and the effective- ness of implicit maximum likelihood estimation,

    A. Mohamed, D. Zhu, W. Vu, M. Elhoseiny, and C. Claudel, “Social- implicit: Rethinking trajectory prediction evaluation and the effective- ness of implicit maximum likelihood estimation,” in Proc. Eur. Conf. Comput. Vis. Springer, 2022, pp. 463–479

    2022

  49. [57]

    Remember intentions: Retrospective-memory-based trajectory prediction,

    C. Xu, W. Mao, W. Zhang, and S. Chen, “Remember intentions: Retrospective-memory-based trajectory prediction,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2022, pp. 6488–6497

    2022

  50. [58]

    Leapfrog diffusion model for stochastic trajectory prediction,

    W. Mao, C. Xu, Q. Zhu, S. Chen, and Y . Wang, “Leapfrog diffusion model for stochastic trajectory prediction,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2023, pp. 5517–5526

    2023

  51. [59]

    Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning,

    C. Xu, R. T. Tan, Y . Tan, S. Chen, Y . G. Wang, X. Wang, and Y . Wang, “Eqmotion: Equivariant multi-agent motion prediction with invariant interaction reasoning,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2023, pp. 1410–1420

    2023

  52. [60]

    Smemo: social memory for trajectory forecasting,

    F. Marchetti, F. Becattini, L. Seidenari, and A. Del Bimbo, “Smemo: social memory for trajectory forecasting,” IEEE Trans. Pattern Anal. Mach. Intell., 2024

    2024

  53. [61]

    Singulartrajectory: Universal trajec- tory predictor using diffusion model,

    I. Bae, Y .-J. Park, and H.-G. Jeon, “Singulartrajectory: Universal trajec- tory predictor using diffusion model,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2024, pp. 17 890–17 901

    2024

  54. [62]

    Higher- order relational reasoning for pedestrian trajectory prediction,

    S. Kim, H.-g. Chi, H. Lim, K. Ramani, J. Kim, and S. Kim, “Higher- order relational reasoning for pedestrian trajectory prediction,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2024, pp. 15 251–15 260

    2024

  55. [63]

    Adaptive trajectory prediction via transferable gnn,

    Y . Xu, L. Wang, Y . Wang, and Y . Fu, “Adaptive trajectory prediction via transferable gnn,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit. , 2022, pp. 6520–6531

    2022

  56. [64]

    Inductive representation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive representation learning on large graphs,” Adv. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  57. [65]

    Skeleton-based action recog- nition using sparse spatio-temporal gcn with edge effective resistance,

    T. Ahmad, L. Jin, L. Lin, and G. Tang, “Skeleton-based action recog- nition using sparse spatio-temporal gcn with edge effective resistance,” Neurocomputing, vol. 423, pp. 389–398, 2021

    2021

  58. [66]

    Hdmixer: Hierarchical dependency with extendable patch for multivariate time series forecasting,

    Q. Huang, L. Shen, R. Zhang, J. Cheng, S. Ding, Z. Zhou, and Y . Wang, “Hdmixer: Hierarchical dependency with extendable patch for multivariate time series forecasting,” in Proc. AAAI Conf. Artif. Intell. , vol. 38, no. 11, 2024, pp. 12 608–12 616

    2024

  59. [67]

    Flashattention: Fast and memory-efficient exact attention with io-awareness,

    T. Dao, D. Fu, S. Ermon, A. Rudra, and C. R ´e, “Flashattention: Fast and memory-efficient exact attention with io-awareness,” Adv. Neural Inf. Process. Syst. , vol. 35, pp. 16 344–16 359, 2022

    2022

  60. [68]

    Efficient transformers: A survey,

    Y . Tay, M. Dehghani, D. Bahri, and D. Metzler, “Efficient transformers: A survey,” ACM Comput. Surv., vol. 55, no. 6, 2022. Ruochen Li is a PhD candidate in the Department of Computer Science at Durham University, UK. He received the BSc and MSc in Computer Science at the Uni...

    2022

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.