Pith. sign in

REVIEW 4 major objections 4 minor 36 references

SEP-GCN: Leveraging Similar Edge Pairs with Temporal and Spatial Contexts for Location-Based Recommender Systems

T0 review · 4 major / 4 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 A recommender that links check-ins sharing time and place beats graph baselines on location data.

desk verdict A clean LightGCN-style extension with a potentially leaking SEP graph construction; the evaluation needs fixing before the claimed gains can be trusted. read the letter →

arxiv 2506.16003 v1 pith:AR5G5TJC submitted 2025-06-19 cs.IR cs.ITmath.IT

Tan Loc Nguyen , Tin T. Tran This is my paper
classification cs.IRcs.ITmath.IT
keywords collaborativefilteringrecommendersystemsedgelearningembeddingpropagationgraphconvolutionalnetworkslocation-basedrecommendationtemporalcontextspatial
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

The paper is trying to establish that a recommender system becomes more accurate and robust when, alongside the usual user-item interaction graph, it builds a second graph whose nodes are individual check-ins and whose edges join check-ins that happen at similar times of the week or in nearby places. SEP-GCN propagates embeddings over this check-in-similarity graph and feeds the result back into user and item embeddings after each convolution step. The authors report consistent wins over strong graph-based baselines on three location datasets, with the largest gains on sparse data and deeper Top-20 lists. If the claim holds, it means the relational context between interactions carries signal that node-level graph recommenders miss.

What carries the argument

The central object is the Similar Edge Pair (SEP) matrix $X$, an $|E|\times|E|$ symmetric sparse matrix whose entry $\sigma(e_1,e_2)$ is nonzero when two check-in edges share at least one weekly time slot and whose strength decays exponentially with great-circle distance relative to the user's median travel distance. The matrix is normalized and used to propagate edge embeddings formed by concatenating user and item embeddings, and the propagated result is blended back into user and item embeddings after each convolution layer.

What would settle it

Rebuild X using only training check-ins and evaluate on a time-ordered split (train on the earliest 70% of check-ins, test on the latest 30%); if SEP-GCN's reported recall@20 improvement over the best comparison baseline, such as 14.89% on Gowalla, does not reproduce, the superiority claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that a location-based recommender improves if each check-in interaction is treated as a node in a secondary graph and linked to other check-ins that fall in the same weekly time slot and lie within a user-relative geographic radius. In SEP-GCN, user and item embeddings from each graph-convolution layer are concatenated into edge embeddings, propagated over that edge-similarity graph, and then folded back into the user and item embeddings with a weighted update. The authors report that this design yields the highest recall, precision, NDCG, and accuracy under Top-5 and Top-20 thresholds on the NYC, Gowalla, and Brightkite datasets, with the largest relative gains on full, sparser versions of the data and on 5-core subsets. They conclude that relational structure between interactions carries predictive signal that node-level graph recommenders underuse.

Load-bearing premise

The load-bearing premise is that the similarity matrix X is built only from training edges and that the random 70/30 split keeps the temporal ordering meaningful; the paper never states that test check-ins are excluded when constructing X.

Editorial extensions

If this is right

  • If the reported gains hold, adding context-similarity edges to a LightGCN-style backbone improves recall@20 by up to 14.89% on full Gowalla and accuracy by up to 7.42% on Brightkite.
  • SEP-GCN keeps improving beyond three propagation layers, where the comparison GCN models degrade, so edge-similarity propagation appears to blunt over-smoothing.
  • Ablations indicate that removing either the spatial-distance filter or the temporal overlap condition lowers performance, so the two signals work jointly rather than independently.
  • Under 5-core sparsity, reported recall@20 improvements over the best baseline reach 23.2% on Gowalla, suggesting particular value near cold-start conditions.

Reading between the lines

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

  • An untested extension is to apply the same edge-pair construction to non-location implicit feedback, using session or device context in place of geographic distance; the paper only evaluates check-in data.
  • Because gains grow with sparsity, combining SEP-style edge links with contrastive learning objectives could push further, though SEP-GCN does not include such a module.
  • A strict temporal split and train-only SEP construction would test whether the reported margins are causal or partly an artifact of test information entering the similarity graph; the authors do not perform this check.
  • The local-time-zone design implies sensitivity to timezone metadata; varying slot granularity (for example, 48 half-hour slots instead of 168 hourly slots) would test whether the temporal overlap mechanism is robust.
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 / 4 minor

Summary. SEP-GCN proposes to augment a LightGCN-style user-item recommender with a secondary graph whose nodes are check-in edges and whose links connect edges with overlapping weekly time slots and small geographic distance. The similarity matrix X is precomputed from the data; edge embeddings are built by concatenating user and item embeddings and propagated over X, then used to update user/item embeddings after each GCN layer. The paper reports experiments on NYC, Gowalla, and Brightkite, claiming consistent improvements over eight baselines across recall, precision, NDCG, and accuracy under top-5 and top-20 thresholds, plus ablations and sparse-data analyses.

Significance. The idea of learning from similar edge pairs rather than from individual nodes or edges is a meaningful conceptual contribution: it lets contextual signals propagate between topologically distant interactions and is a natural fit for location-based recommendation. The use of user-local time zones and a personalized median-distance decay in the similarity definition is thoughtful. The ablation study in Table 5 and the sparse-data experiments in Table 6 address relevant questions. However, the empirical support for the central claim is currently undermined by a potentially serious data-leakage issue in the construction of the SEP graph, the use of a random temporal split for a time-aware model, and incomplete statistical reporting. If the leakage concern is confirmed, the reported margins over LightGCN would be substantially inflated because the SEP module is the only architectural difference. The paper also does not provide code or exact hyperparameter values for the new weights, which prevents independent auditing.

major comments (4)
  1. [Section 3.2 (Algorithm 1), Definition 2, Section 5.1.1] The paper never states that the SEP matrix X is constructed from the training split only. Algorithm 1 takes 'E: Edge set' as input with no filtering, Definition 2 defines G=(V,E,X) on the full check-in graph, and Section 5.1.1 only says the data were split 70/30. If E includes test edges, then each held-out check-in appears as a node in the SEP graph, and its timestamp and location are used in Equations (8)-(9) while fitting parameters. Since the SEP module is the only difference from LightGCN, the reported gains in Table 2 (e.g., +14.89% recall@20 on Gowalla) could be an artifact of test-label leakage. The authors must specify explicitly that X is computed from training check-ins only, and that the median-distance normalization in Equation (2) is also computed on training distances only.
  2. [Section 5.1.1 and Definition 5] Even if X were restricted to training edges, the random 70/30 split is inappropriate for a time-aware model. The model explicitly uses weekly time-slot overlap, so training interactions that occur after test interactions in the same user's history would allow the model to use future time-slot co-occurrences when predicting earlier behavior. This temporal inversion is a correctness risk for the time-aware framing of SEP-GCN. The authors should use a temporal split (e.g., leave-last-out per user or a global timestamp cutoff) and re-run the comparisons.
  3. [Section 5.1.3 and Table 2] The statistical claims are not supported by the reported numbers. Section 5.1.3 says results are averaged over 5 runs and significant values are marked at p<0.01, but Table 2 reports only point estimates with no standard deviations, no standard errors, and no exact p-values. The 'Accuracy' metric is also undefined: Section 5.1.3 defines it as 'the proportion of correct predictions over all predictions,' which is not a well-defined notion for top-k ranking; the tables report accuracy@5 and accuracy@20 without a formal definition. The authors should define accuracy precisely and report means with confidence intervals or at least standard deviations for all baselines and for SEP-GCN.
  4. [Equations (8)-(9), Algorithm 2, Section 5.1.4] The update weights are presented inconsistently and are not specified. Equations (8) and (9) use alpha and beta as the weights for user and item embeddings, while Algorithm 2 uses a single gamma for the same update. Section 5.1.4 lists the learning rate, L2 coefficient, number of layers, and embedding size, but never gives the values of alpha, beta, or gamma, nor any sensitivity analysis. Since these weights control the influence of the proposed SEP module, their values are needed to assess whether the reported improvements come from the SEP mechanism itself or from particular settings of the new hyperparameters.
minor comments (4)
  1. [Section 3.1, Definition 2] Definition 2 defines E using 'R_{i,j}=1' but writes edges as e_{u,i}; the subscripts are inconsistent and should be aligned (e.g., R_{u,i}).
  2. [Section 5.3.4] The heading 'The effectiveness on spare dataset' contains a typo; it should read 'sparse dataset'.
  3. [Section 2.1 and References] The introduction cites 'NGCF [10]' but the reference list entry for NGCF is [24]; please correct the citation numbering and make the related-work references consistent.
  4. [Section 4.2.1] Equation (6) uses superscripts k and l for the same layer in adjacent equations; please use one consistent superscript for the propagation layer to avoid confusion.

Circularity Check

0 steps flagged · score 2.0 of 10

No derivational circularity found; the only circularity-adjacent element is a non-load-bearing self-citation to the corresponding author's prior GCN-LOC work.

full rationale

Walking the derivation chain, no predicted quantity reduces to a fitted or definitionally equivalent input. The SEP matrix X is precomputed (Algorithm 1, Definition 6) from timestamps and locations; edge embeddings are concatenations of user/item embeddings (Eq. 6); propagation (Eqs. 7-9) is a standard coupled message-passing update; and BPR training fits the embeddings. Nothing in Eqs. (8)-(12) defines a prediction in terms of its own target. The one self-citation, GCN-LOC [21] (Section 2.2 and baseline list in 5.1.2), is prior work by the corresponding author, but it is used only as a comparative baseline and related-work context and is not load-bearing for the SEP-GCN derivation. A separate concern, not circularity, is that Section 5.1.1's 70/30 random split and Algorithm 1's use of the full edge set E never explicitly exclude test check-ins from the SEP graph; if test edges enter X, the reported gains would be inflated by label leakage. This should be fixed by stating that X is built from training edges only and by using a time-respecting split, but it does not make the derivation circular.

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

The central empirical claim rests on the precomputed SEP similarity matrix X. Its construction depends on the 168-slot temporal discretization, a location-versus-edge time slot ambiguity, an uncalibrated distance decay alpha, and a data split that must not leak test edges into X. The paper reports none of the values for alpha, beta, or gamma and does not specify how X is split. These are the main cost drivers of the contribution.

free parameters (5)
  • alpha (median-distance similarity score) = not reported
    Equation (2) uses alpha as the similarity value at the median distance. The paper does not give its value or a sensitivity analysis.
  • alpha (user embedding update weight) = not reported
    Equation (8) uses alpha to balance the propagated user embedding against the SEP update. The same symbol alpha is reused from Equation (2), and no value is reported.
  • beta (item embedding update weight) = not reported
    Equation (9) uses beta to balance the propagated item embedding against the SEP update. No value is reported.
  • gamma (SEP aggregation weight in Algorithm 2) = not reported
    Algorithm 2 uses gamma in [0,1] to blend SEP-aggregated embeddings with base embeddings. Its relationship to alpha and beta is unclear, and no value is reported.
  • Weekly time slot granularity (7 days x 24 hours) = 168
    Definition 3 chooses this discretization. Results may depend on this granularity, but no ablation over it is reported.
assumptions (5)
  • domain assumption Weekly time is discretized into 168 one-hour slots, and check-ins sharing a slot are treated as temporally similar.
    Definition 3 and Figure 1 introduce this granularity. The choice of 168 slots and the equivalence of slot overlap with behavioral similarity are asserted rather than derived.
  • ad hoc to paper The time slot set defined for a location can also be attached to each user-item edge.
    Definition 3 defines time slot sets for locations, but Definition 5 applies intersection to edge time slot sets t1 and t2. The paper never specifies how an individual edge obtains its time slot set.
  • ad hoc to paper Distance similarity follows sigma = exp((d/median) * ln(alpha)), with the median distance acting as a personalized radius filter.
    Equation (2) introduces this functional form for this paper. The value of alpha and the use of a global or per-user median are not specified consistently.
  • domain assumption A random 70/30 split is a valid evaluation setup for a time-aware recommender system.
    Section 5.1.1 states the split. Random splitting can place test check-ins earlier than training check-ins, which contradicts the temporal modeling premise.
  • domain assumption Local time zones are known for every check-in and are used to assign time slots.
    Definition 5 requires user-local time zones. The datasets' timezone coverage and how missing timezones are handled are not described.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SEP-GCN: Leveraging Similar Edge Pairs with Temporal and Spatial Contexts for Location-Based Recommender Systems." pith.science (2026). https://pith.science/paper/AR5G5TJC

@misc{pith2026250616003,
  author       = {Pith},
  title        = {Pith review of: SEP-GCN: Leveraging Similar Edge Pairs with Temporal and Spatial Contexts for Location-Based Recommender Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AR5G5TJC}},
  note         = {Machine review of arXiv:2506.16003}
}
read the original abstract

Recommender systems play a crucial role in enabling personalized content delivery amidst the challenges of information overload and human mobility. Although conventional methods often rely on interaction matrices or graph-based retrieval, recent approaches have sought to exploit contextual signals such as time and location. However, most existing models focus on node-level representation or isolated edge attributes, underutilizing the relational structure between interactions. We propose SEP-GCN, a novel graph-based recommendation framework that learns from pairs of contextually similar interaction edges, each representing a user-item check-in event. By identifying edge pairs that occur within similar temporal windows or geographic proximity, SEP-GCN augments the user-item graph with contextual similarity links. These links bridge distant but semantically related interactions, enabling improved long-range information propagation. The enriched graph is processed via an edge-aware convolutional mechanism that integrates contextual similarity into the message-passing process. This allows SEP-GCN to model user preferences more accurately and robustly, especially in sparse or dynamic environments. Experiments on benchmark data sets show that SEP-GCN consistently outperforms strong baselines in both predictive accuracy and robustness.

Figures

Figures reproduced from arXiv: 2506.16003 by the authors.

Figure 1
Figure 1. Weekly Time Slots and Time Slot Sets of Check-ins. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. The proposed SEP-GCN model architecture, incorporating spatiotemporal edge similarity into GCN-based propagation [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Comparison of Recall@20 and NDCG@20 for dif [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 14 canonical work pages

  1. [1]

    Edge-featured Graph Neural Architecture Search

    Y. Cai and F. Wu. 2021. Edge-featured graph neural architecture search. In Proceedings of NeurIPS. https://arxiv.org/abs/2109.01356

    work page Pith review arXiv 2021

  2. [2]

    Jinpeng Chen, Wen Zhang, Pei Zhang, Pinguang Ying, Kun Niu, and Ming Zou

    work page

  3. [3]

    Chen and J

    L. Chen and J. Chen. 2021. Edge-featured graph attention network for improved learning.arXiv preprint arXiv:2101.07671. https://arxiv.org/abs/2101.07671

    arXiv 2021

  4. [5]

    Jie Feng, Yong Li, Chao Zhang, Funing Sun, Fanchao Meng, Ang Guo, and Depeng Jin. 2018. Deepmove: predicting human mobility with attentional recurrent networks. In (Apr. 2018), 1459–1468. doi:10.1145/3178876.3186058

    arXiv 2018

  5. [6]

    M Fey and J Lenssen. 2023. Pytorch geometric documentation. (2023). https://p ytorch-geometric.readthedocs.io

    work page 2023

  6. [7]

    Gong and H

    M. Gong and H. Cheng. 2019. Exploiting edge features for graph neural net- works. InProceedings of CVPR. https://openaccess.thecvf.com/content_CVPR_2 019/papers/Gong_Exploiting_Edge_Features_for_Graph_Neural_Networks _CVPR_2019_paper.pdf

    work page 2019

  7. [8]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representa- tion learning on large graphs. InProceedings of the 31st International Conference on Neural Information Processing Systems(NIPS’17). Curran Associates Inc., Long Beach, California, USA, 1025–1035.isbn: 9781510860964. doi:10.5555/329 4771.3294869. arXiv Preprint, June 2025, Nguye...

    work page arXiv 2017

  8. [9]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. ACM, 639–648. doi:10.11 45/3397271.3401063

    arXiv 2020

Show all 36 references
  1. [10]

    Xiangnan He, Lizi Liao, Hanwang Zhang, Liqiang Nie, Xia Hu, and Tat-Seng Chua. 2017. Neural collaborative filtering. InProceedings of the 26th Interna- tional Conference on World Wide Web. International World Wide Web Confer- ences Steering Committee, 173–182. https://doi.org/...

    2017

  2. [11]

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk

  3. [12]

    Tianhao Huang, Xuan Pan, Xiangrui Cai, Ying Zhang, and Xiaojie Yuan. 2024. Learning time slot preferences via mobility tree for next poi recommendation. Proceedings of the AAAI Conference on Artificial Intelligence, 38, 8, (Mar. 2024), 8535–8543. doi:10.1609/aaai.v38i8.28697

    2024 doi

  4. [13]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-supervised classification with graph convolutional networks. In5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. https://openreview.ne...

    2017

  5. [14]

    Yehuda Koren. 2010. Collaborative filtering with temporal dynamics.Commun. ACM, 53, 4, (Apr. 2010), 89–97. doi:10.1145/1721654.1721677

    2010

  6. [15]

    Srijan Kumar, Xikun Zhang, and Jure Leskovec. 2019. Predicting dynamic embedding trajectory in temporal interaction networks. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining(KDD ’19). Association for Computing Machinery, Ancho...

    2019

  7. [16]

    Yuen Hoi Lau and Raymond Chi-Wing Wong. 2021. Spatio-temporal graph convolutional networks for traffic forecasting: spatial layers first or temporal layers first? InProceedings of the 29th International Conference on Advances in Geographic Information Systems(SIGSPATIAL ’21). ...

    2021 doi

  8. [17]

    Hao Li, Peng Yue, Shangcheng Li, Chenxiao Zhang, and Can Yang. 2024. Spatio- temporal intention learning for recommendation of next point-of-interest. Geo-spatial Information Science, 27, 2, 384–397. eprint: https://doi.org/10.1080 /10095020.2023.2179428. doi:10.1080/10095020....

    2024

  9. [18]

    Zihan Lin, Changxin Tian, Yupeng Hou, and Wayne Xin Zhao. 2022. Improving graph collaborative filtering with neighborhood-enriched contrastive learning. Proceedings of the ACM Web Conference 2022. https://api.semanticscholar.org /CorpusID:246823233

    2022

  10. [19]

    Yingtao Luo, Qiang Liu, and Zhaocheng Liu. 2021. Stan: spatio-temporal at- tention network for next location recommendation. InProceedings of the Web Conference 2021(WWW ’21). Association for Computing Machinery, Ljubljana, Slovenia, 2177–2185.isbn: 9781450383127. doi:10.1145/...

    2021

  11. [20]

    Steffen Rendle, Christoph Freudenthaler, Zeno Gantner, and Lars Schmidt- Thieme. 2009. Bpr: bayesian personalized ranking from implicit feedback.ArXiv, abs/1205.2618. https://api.semanticscholar.org/CorpusID:10795036

    2009 arXiv

  12. [21]

    Tran, Vaclav Snasel, and Thuan Q

    Tin T. Tran, Vaclav Snasel, and Thuan Q. Nguyen. 2024. Collaborative filtering by graph convolution network in location-based recommendation system.KSII Transactions on Internet and Information Systems, 18, 7, (July 2024), 1868–1887. doi:10.3837/tiis.2024.07.008

    2024 doi

  13. [22]

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. 2018. Graph attention networks. InInternational Con- ference on Learning Representations (ICLR). doi:10.48550/arXiv.1710.10903

    2018 doi arXiv

  14. [23]

    Hongwei Wang, Fuzheng Zhang, Miao Zhao, Wenjie Li, Xing Xie, and Minyi Guo. 2019. Multi-task feature learning for knowledge graph enhanced recom- mendation. InThe World Wide Web Conference(WWW ’19). Association for Computing Machinery, San Francisco, CA, USA, 2000–2010.isbn: 9...

    2019

  15. [24]

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. 2019. Neural graph collaborative filtering. InProceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR’19). Association for Computing Machinery, Pari...

    2019

  16. [26]

    Xinfeng Wang, Fumiyo Fukumoto, Jin Cui, Yoshimi Suzuki, Jiyi Li, and Dongjin Yu. 2023. Eedn: enhanced encoder-decoder network with local and global con- text learning for poi recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Developme...

    2023

  17. [27]

    Shu Wu, Yuyuan Tang, Yanqiao Zhu, Liang Wang, Xing Xie, and Tieniu Tan

  18. [28]

    Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong Long, Chengqi Zhang, and Philip S. Yu. 2021. A comprehensive survey on graph neural networks.IEEE Transactions on Neural Networks and Learning Systems, 32, 1, 4–24. doi:10.1109 /TNNLS.2020.2978386

    2021

  19. [29]

    Da Xu, Chuanwei Ruan, Evren Korpeoglu, Sushant Kumar, and Kannan Achan

  20. [30]

    Reza Yeganegi, Saman Haratizadeh, and Morteza Ebrahimi. 2024. Star: a session- based time-aware recommender system.Neurocomputing, 573, 127104. doi:http s://doi.org/10.1016/j.neucom.2023.127104

    2024

  21. [31]

    Muhan Zhang and Yixin Chen. 2018. Link prediction based on graph neural networks. InAdvances in Neural Information Processing Systems. S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, (Eds.) Vol. 31. Curran Associates, Inc. https://proceedings...

    2018

  22. [33]

    Yi Zhang, Lei Sang, and Yiwen Zhang. 2024. Exploring the individuality and collectivity of intents behind interactions for graph collaborative filtering. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval(SIGIR ’2...

    2024

  23. [34]

    Ziwei Zhang, Peng Cui, and Wenwu Zhu. 2022. Deep learning on graphs: a survey.IEEE Trans. on Knowl. and Data Eng., 34, 1, (Jan. 2022), 249–270. doi:10.1109/TKDE.2020.2981333

    2022

  24. [35]

    Yuchen Zhou, Hongtao Huo, Zhiwen Hou, Lingbin Bu, Jingyi Mao, Yifan Wang, Xiaojun Lv, and Fanliang Bu. 2023. Co-embedding of edges and nodes with deep graph convolutional neural networks.Scientific Reports, 13, 1, (Oct. 2023), 16966. doi:10.1038/s41598-023-44224-1

    2023 doi

  25. [2015]

    https://api.semanticscholar.org/CorpusID:260446846

    Session-based recommendations with recurrent neural networks.CoRR, abs/1511.06939. https://api.semanticscholar.org/CorpusID:260446846

    arXiv

  26. [2018]

    Complexity, 2018, 1, 6928605

    Exploiting spatial and temporal for point of interest recommendation. Complexity, 2018, 1, 6928605. eprint: https://onlinelibrary.wiley.com/doi/pdf/1 0.1155/2018/6928605. doi:https://doi.org/10.1155/2018/6928605

    2018 doi

  27. [2019]

    Session-based recommendation with graph neural networks. InProceed- ings of the Thirty-Third AAAI Conference on Artificial Intelligence and Thirty-First Innovative Applications of Artificial Intelligence Conference and Ninth AAAI Sym- posium on Educational Advances in Artifici...

    doi

  28. [2020]

    Inductive representation learning on temporal graphs. (Feb. 2020). doi:10 .48550/arXiv.2002.07962

    2020 doi arXiv

Pith tools

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