Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Spatial-Temporal Knowledge Distillation for Takeaway Recommendation

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read STKDRec, a two-stage teacher-student model, outperforms nine takeaway recommenders on three real-world city datasets by distilling spatial-temporal knowledge from a knowledge graph into a lightweight Transformer.

desk verdict A clean two-stage distillation recipe for spatial-temporal recommendation that is one clarifying sentence away from being trustworthy. read the letter →

arxiv 2412.16502 v2 pith:AZLFR6FH submitted 2024-12-21 cs.LG cs.IR

Shuyuan Zhao , Wei Chen , Boyan Shi , Liyong Zhou , Shuohao Lin , Huaiyu Wan This is my paper · ORCID
classification cs.LGcs.IR
keywords takeawayrecommendationsequentialknowledgedistillationgraphspatial-temporalmodelingTransformerneuralnetworksgeohash
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 tries to show that a takeaway recommender can combine the relational structure of a knowledge graph with the sequential logic of order histories, without paying the full inference cost of graph neural networks. It proposes STKDRec, which pre-trains a spatial-temporal knowledge graph encoder as a teacher and distills its knowledge into a lighter spatial-temporal Transformer that reads purchase sequences. Across the Wuhan, Sanya, and Taiyuan datasets, STKDRec ranks first on every reported metric, for example raising Wuhan HR@10 from 0.7976 (best baseline) to 0.8229. The spatial position embeddings for region and distance, plus the distillation loss, carry the performance gain.

What carries the argument

The machinery is a two-stage teacher-student setup. The teacher is a GNN-based STKG encoder that aggregates messages over a sampled subgraph of user-takeaway-time-distance-attribute triples, producing user-specific gated representations and soft label distributions over all takeaways. The student is a spatial-temporal Transformer that adds a learnable spatial position embedding (a linear combination of embeddings for the geohash region and the spherical distance of each purchased item) to the usual token and absolute position embeddings, then applies masked self-attention. The two are joined by a distillation loss that minimizes the KL divergence between the teacher's softened soft labels and the student's predictions, $\mathcal{L}_{KD} = \mathrm{KL}(Y'_x/\tau \,\|\, \hat{Y}_x/\tau)$, weighted by a coefficient $\alpha$ against the supervised cross-entropy loss. This carries graph-level collaborative and spatial-temporal knowledge into a lightweight sequential model without running GNN inference at serving time.

What would settle it

Rebuild the STKG after removing, for every user, all triples that mention their last purchased takeaway, retrain the teacher and student, and re-evaluate. If HR@10 and NDCG@10 for STKDRec on Wuhan, Sanya, and Taiyuan drop to within noise of the best baseline, the original result depended on leakage. Additionally, check the teacher's top-k soft-label probabilities for test items: high mass on a user's held-out purchase before any training update would be direct evidence of contamination.

Watch

Extended reading notes

Core claim

STKDRec's central claim is that spatial-temporal knowledge can be transferred from a graph to a sequence model through distillation, yielding better takeaway recommendations than either representation alone. The paper constructs a spatial-temporal knowledge graph whose edges are user-takeaway purchases with time, distance, and attribute relations, and pre-trains a GNN encoder on sampled subgraphs. A student Transformer then models the user's purchase sequence with learnable spatial position embeddings built from geohash regions and spherical distances, and is trained to imitate the teacher's soft label distribution in addition to the true next-item labels. On the three city datasets the full model outperforms all nine baselines on all eighteen HR@k and NDCG@k cells, with notable margins over the strongest baseline BSARec.

Load-bearing premise

The load-bearing premise is that the spatial-temporal knowledge graph contains no information about the held-out last purchase of any user; if test purchases appear in the graph's triples, the teacher's soft labels can leak the answer into training, and the reported gains would not reflect genuine learning.

Editorial extensions

If this is right

  • If STKDRec's gains hold, a food-delivery platform can serve a Transformer-level recommender at inference time with the accuracy of a graph model, since the teacher runs only during training.
  • The method's spatial position embedding suggests a general recipe for injecting continuous geographic variables into sequential recommenders, with applications in ride-hailing, delivery routing, and visit prediction.
  • The 0.03 to 0.05 absolute improvements in HR@10 over the best baseline, on dense city datasets, indicate that spatial-temporal context carries information that pure sequence models miss.
  • The parameter sensitivity tests show that performance is robust to the number of sampled neighbors but degrades with large distillation temperature, so practitioners can fix the sampling size while tuning only the temperature.

Reading between the lines

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

  • A check the paper does not run: the data split marks each user's last purchase as test, but the STKG construction section never states that test purchases are excluded from the graph triples; if they are included, the teacher's soft labels could leak the answer into the distillation loss, inflating the reported margins.
  • A natural extension is to apply the same distillation scheme to other knowledge-rich side information, such as review text or merchant attributes, where the teacher and student have different inductive biases.
  • The teacher-student asymmetry also points to a compression direction: the student could be made smaller than the tested 256-dimension setting, trading a little accuracy for larger serving-time savings, a trade-off the paper does not explore.
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 / 5 minor

Summary. The paper proposes STKDRec, a two-stage takeaway recommendation model. In the first stage, a spatial-temporal knowledge graph (STKG) encoder is pre-trained to capture high-order spatial-temporal and collaborative associations from triples such as (user, time, item) and (user, distance, item). In the second stage, a spatial-temporal Transformer (ST-Transformer) is trained as a student, using a spatial-enhanced sequence representation that combines region and distance embeddings, and it is supervised by both ground-truth next-item labels and a KL distillation loss against the teacher's soft labels. The method is evaluated on three Ele.me datasets (Wuhan, Sanya, Taiyuan) against nine baselines, with HR@k and NDCG@k metrics, plus ablations and a parameter sensitivity study. Code is released.

Significance. If the empirical results are not confounded by label leakage, the paper makes a practical contribution by combining graph-based and sequence-based spatial-temporal modeling through knowledge distillation, with a plausible computational-efficiency argument and broad evaluation across three datasets and nine baselines. The ablations are helpful in isolating the contributions of the spatial position embedding and the distillation loss, and the released code is a strength. However, the headline claim that STKDRec 'significantly outperforms' state-of-the-art baselines is currently not well supported because of a potentially load-bearing data-leakage issue in the STKG construction and because the experimental protocol lacks statistical rigor and omits the closest spatial-temporal baselines.

major comments (4)
  1. [Spatial-Temporal Knowledge Graph Encoder; Eq. (4), (10)-(13)] The data-split description in the Datasets paragraph states only that the last purchased takeaway per user is test data, the second-to-last is validation, and the rest is training; it never states that the STKG is constructed exclusively from the training portion. The STKG contains triples (u, time, v) and (u, distance, v), and the subgraph sampling procedure treats each sequence item as a center and recursively samples neighbors to depth m=2. Because the user node is adjacent to every item the user purchased, a depth-2 path from any context item to the target item passes through the user. If the STKG includes held-out purchases, the teacher's soft labels in Eq. (4) can assign high probability to the test item, and the distillation loss in Eq. (11) can leak that answer into the student; if the STKG excludes held-out purchases but still contains training-time purchase edges, Eq. (10) still allows the teacher to see the answer to each training prediction. The authors must specify how the STKG is split and should remove the user-to-target edge for the item being predicted. As written, the Table 2 margins cannot be cleanly attributed to learned spatial-temporal knowledge.
  2. [Experimental Results; Table 2] The abstract and the Experimental Results section use the phrase 'significantly outperforms', but Table 2 reports a single run with no error bars, standard deviations, or significance tests. Several margins are small enough that they could be within seed-to-seed variation; for example, Sanya HR@5 is 0.8770 for STKDRec versus 0.8623 for BSARec, and Taiyuan HR@10 is 0.8789 versus 0.8661. The authors should report mean and standard deviation over multiple random seeds and, if 'significant' is claimed, a paired statistical test.
  3. [Related Work; Table 2] The Related Work discusses StEN (Lin et al. 2022) and BASM (Du et al. 2023a) as spatial-temporal methods for online food ordering, yet neither method appears in the baseline comparison in Table 2. Since the paper's central claim is improved spatial-temporal takeaway recommendation, omitting these closest spatial-temporal baselines leaves the 'state-of-the-art' claim incomplete. The authors should add these methods or explicitly justify why they are not applicable to this sequential recommendation setting.
  4. [Experimental Setting; Metrics] The Metrics paragraph says that evaluation pairs the actual test takeaway with 100 randomly sampled negative takeaways and ranks them. This candidate-sampling protocol yields different absolute HR/NDCG values than full-corpus ranking and can favor models that rank popular items highly. The text does not explicitly state that this protocol is applied identically to every baseline method, so the comparisons in Table 2 may not reflect the top-k recommendation task as defined in the Problem Formulation. The authors should confirm that all methods use the same sampling procedure and, ideally, also report full-corpus metrics.
minor comments (5)
  1. [Eq. (10)-(12)] The ground-truth label vector Y_x is used in the pre-training, distillation, and recommendation losses but is never formally defined; the authors should state that it is the one-hot encoding of the next purchased takeaway in the training sequence.
  2. [Figure 3] Figure 3 uses 'Recall@10' on the y-axes, while the metrics are defined as HR@k (Hit Rate) in the Experimental Setting; these names should be made consistent.
  3. [Ablation Studies; Table 3] The ablation discussion refers to a variant '-w/o KD+SP', while Table 3 labels the same variant '-w/o SP+KD'; the notation should be unified.
  4. [Table 3; Table 2] The '-w/o SP+KD' row in Table 3 is numerically identical to the SASRec row in Table 2 for all datasets (e.g., Wuhan HR@10=0.7796 and NDCG@10=0.6911). The authors should clarify whether this variant is exactly SASRec and, if so, why it is not listed among the baselines in Table 2.
  5. [Figure 2] The label 'Absolute Positon Embedding' in Figure 2 contains a typo and should read 'Absolute Position Embedding'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the teacher–student distillation chain is internally consistent; the main risk is a possible data-leakage gap in STKG construction, not a circular reduction.

full rationale

The central empirical claim (Table 2) is a benchmark comparison, not a derivation, and the model architecture is internally consistent: the STKG encoder is pre-trained with the cross-entropy objective (Eq. 10), its soft labels (Eq. 4) are distilled into the ST-Transformer via KL divergence (Eq. 11) combined with a supervised loss (Eq. 12). None of these equations re-introduces the target as a fitted constant or as an input definition. The self-citations (Chen et al. 2022, 2024; Lin et al. 2022) appear only in related-work motivation and do not carry the load of the claimed improvement. No uniqueness theorem or ansatz is imported from the authors' prior work. The one substantive concern is a possible soundness gap, not circularity: the data-preparation paragraph designates the last purchase as test and the second-to-last as validation, but the STKG construction section does not explicitly state that test/validation purchase edges are excluded from the triples (u, time, v) and (u, distance, v). If they are not excluded, the teacher's soft labels could leak the answer through Eq. 11. This is a data-leakage/correctness risk that should be verified, but because the paper's text does not exhibit the inclusion, it cannot be scored as a demonstrated circular reduction. Verdict: no significant circularity.

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

The model relies on standard deep learning building blocks and on the availability of rich geospatial and temporal features. The main unstated assumption is that the knowledge graph built for pre-training does not include the test purchase. Hyperparameters listed above are tuned on validation and are part of the model design rather than derived constants.

free parameters (5)
  • distillation weight alpha = 0.2
    Chosen to balance KL divergence and cross-entropy in Eq. (13); fixed after validation.
  • distillation temperature tau = 7 on Sanya and Taiyuan, best of {1,3,5,7,9}
    Controls softness of teacher labels in Eq. (11); the parameter study shows performance declines beyond 7.
  • neighbor sample size s = [20,20] on Sanya, [10,10] on Taiyuan, from {[5,5],[10,10],[15,15],[20,20]}
    Number of sampled neighbors in STKG subgraph; chosen on validation data.
  • maximum sequence length n = 128
    Truncation and padding length for purchase sequences; also used as batch size.
  • embedding dimension d = 256
    Latent dimension for entities, tokens, and spatial embeddings; chosen in implementation details.
assumptions (5)
  • standard math Attention, GNN message passing, cross-entropy, and KL divergence are used as given, with no proof sketches needed.
    Equations (1)-(9) and (11) rest on standard definitions from the cited literature.
  • domain assumption The next purchase is modeled as P(v_{|x|+1} | x) from the chronological purchase sequence alone.
    Problem formulation in the Preliminaries section.
  • domain assumption Spatial-temporal relations among users, foods, time, distance, and attributes can be encoded as triples in a single knowledge graph.
    STKG construction in Section 3.1.
  • ad hoc to paper The teacher's soft labels carry knowledge that improves the student beyond ground-truth labels.
    Core premise of the STKD loss in Eq. (11); without it the distillation term is noise.
  • ad hoc to paper The STKG used for pre-training excludes the held-out final purchase of each user.
    Required for a clean evaluation, but never stated in the data processing paragraph.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Spatial-Temporal Knowledge Distillation for Takeaway Recommendation." pith.science (2026). https://pith.science/paper/AZLFR6FH

@misc{pith2026241216502,
  author       = {Pith},
  title        = {Pith review of: Spatial-Temporal Knowledge Distillation for Takeaway Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AZLFR6FH}},
  note         = {Machine review of arXiv:2412.16502}
}
read the original abstract

The takeaway recommendation system aims to recommend users' future takeaway purchases based on their historical purchase behaviors, thereby improving user satisfaction and boosting merchant sales. Existing methods focus on incorporating auxiliary information or leveraging knowledge graphs to alleviate the sparsity issue of user purchase sequences. However, two main challenges limit the performance of these approaches: (1) capturing dynamic user preferences on complex geospatial information and (2) efficiently integrating spatial-temporal knowledge from both graphs and sequence data with low computational costs. In this paper, we propose a novel spatial-temporal knowledge distillation model for takeaway recommendation (STKDRec) based on the two-stage training process. Specifically, during the first pre-training stage, a spatial-temporal knowledge graph (STKG) encoder is trained to extract high-order spatial-temporal dependencies and collaborative associations from the STKG. During the second spatial-temporal knowledge distillation (STKD) stage, a spatial-temporal Transformer (ST-Transformer) is employed to comprehensively model dynamic user preferences on various types of fine-grained geospatial information from a sequential perspective. Furthermore, the STKD strategy is introduced to transfer graph-based spatial-temporal knowledge to the ST-Transformer, facilitating the adaptive fusion of rich knowledge derived from both the STKG and sequence data while reducing computational overhead. Extensive experiments on three real-world datasets show that STKDRec significantly outperforms the state-of-the-art baselines.

Figures

Figures reproduced from arXiv: 2412.16502 by the authors.

Figure 1
Figure 1. An illustrative example highlighting the impor [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall architecture of STKDRec, consisting of two stages: the pre-training stage and the STKD stage. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Study on different knowledge fusion methods. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Study on different temperature τ and the number of neighbor nodes sampled s. Parameter Sensitivity Study We conduct experiments on the Sanya and Taiyuan datasets to investigate the impact of two hyper-parameters: the num￾ber of neighbor nodes sampled s in the STKG and …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 17 canonical work pages

  1. [1]

    Chen, W.; Wan, H.; Guo, S.; Huang, H.; Zheng, S.; Li, J.; Lin, S.; and Lin, Y. 2022. Building and exploiting spatial--temporal knowledge graph for next POI recommendation. Knowledge-Based Systems, 258: 109951

    work page 2022

  2. [2]

    Chen, W.; Wan, H.; Wu, Y.; Zhao, S.; Cheng, J.; Li, Y.; and Lin, Y. 2024. Local-global history-aware contrastive learning for temporal knowledge graph reasoning. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), 733--746. IEEE

    work page 2024

  3. [3]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2018. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805

    arXiv 2018

  4. [4]

    Du, B.; Lin, S.; Gao, J.; Ji, X.; Wang, M.; Zhou, T.; He, H.; Jia, J.; and Hu, N. 2023 a . BASM: A bottom-up adaptive spatiotemporal model for online food ordering service. In 2023 IEEE 39th International Conference on Data Engineering (ICDE), 3549--3562. IEEE

    work page 2023

  5. [5]

    Du, X.; Yuan, H.; Zhao, P.; Qu, J.; Zhuang, F.; Liu, G.; Liu, Y.; and Sheng, V. S. 2023 b . Frequency enhanced hybrid attention network for sequential recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 78--88

    work page 2023

  6. [6]

    Gao, Y.; He, Y.; Kan, Z.; Han, Y.; Qiao, L.; and Li, D. 2023. Learning joint structural and temporal contextualized knowledge embeddings for temporal knowledge graph completion. In Findings of the Association for Computational Linguistics: ACL 2023, 417--430

    work page 2023

  7. [7]

    Hamilton, W.; Ying, Z.; and Leskovec, J. 2017. Inductive representation learning on large graphs. Advances in neural information processing systems, 30

    2017

  8. [8]

    Hidasi, B.; Karatzoglou, A.; Baltrunas, L.; and Tikk, D. 2015. Session-based recommendations with recurrent neural networks. arXiv preprint arXiv:1511.06939

    arXiv 2015

Show all 29 references
  1. [9]

    Hinton, G.; Vinyals, O.; and Dean, J. 2015. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531

    2015 arXiv

  2. [10]

    Kang, S.; Hwang, J.; Kweon, W.; and Yu, H. 2021. Topology distillation for recommender system. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, 829--839

    2021

  3. [11]

    Kang, S.; Kweon, W.; Lee, D.; Lian, J.; Xie, X.; and Yu, H. 2023. Distillation from heterogeneous models for top-k recommendation. In Proceedings of the ACM Web Conference 2023, 801--811

    2023

  4. [12]

    Kang, W.-C.; and McAuley, J. 2018. Self-attentive sequential recommendation. In 2018 IEEE international conference on data mining (ICDM), 197--206. IEEE

    2018

  5. [13]

    Lin, S.; Yu, Y.; Ji, X.; Zhou, T.; He, H.; Sang, Z.; Jia, J.; Cao, G.; and Hu, N. 2022. Spatiotemporal-enhanced network for click-through rate prediction in location-based services. arXiv preprint arXiv:2209.09427

    2022 arXiv

  6. [14]

    Liu, H.; Zhu, Y.; and Wu, Z. 2023. Knowledge graph-based behavior denoising and preference learning for sequential recommendation. IEEE Transactions on Knowledge and Data Engineering

    2023

  7. [15]

    Qiu, R.; Huang, Z.; Yin, H.; and Wang, Z. 2022. Contrastive learning for representation degeneration problem in sequential recommendation. In Proceedings of the fifteenth ACM international conference on web search and data mining, 813--823

    2022

  8. [16]

    Rendle, S.; Freudenthaler, C.; and Schmidt-Thieme, L. 2010. Factorizing personalized markov chains for next-basket recommendation. In Proceedings of the 19th international conference on World wide web, 811--820

    2010

  9. [17]

    Shi, L.; Yang, J.; Lv, P.; Yuan, L.; Kou, F.; Luo, J.; and Xu, M. 2024. Self-derived knowledge graph contrastive learning for recommendation. In Proceedings of the 32nd ACM International Conference on Multimedia, 7571--7580

    2024

  10. [18]

    Shin, Y.; Choi, J.; Wi, H.; and Park, N. 2024. An attentive inductive bias for sequential recommendation beyond the self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, 8984--8992

    2024

  11. [19]

    Sun, F.; Liu, J.; Wu, J.; Pei, C.; Lin, X.; Ou, W.; and Jiang, P. 2019. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In Proceedings of the 28th ACM international conference on information and knowledge management, 1441--1450

    2019

  12. [20]

    Tang, J.; and Wang, K. 2018. Personalized top-n sequential recommendation via convolutional sequence embedding. In Proceedings of the eleventh ACM international conference on web search and data mining, 565--573

    2018

  13. [21]

    Wu, S.; Wan, H.; Chen, W.; Wu, Y.; Shen, J.; and Lin, Y. 2023. Towards enhancing relational rules for knowledge graph link prediction. arXiv preprint arXiv:2310.13411

    2023 arXiv

  14. [22]

    Xia, X.; Yin, H.; Yu, J.; Wang, Q.; Xu, G.; and Nguyen, Q. V. H. 2022. On-device next-item recommendation with self-supervised knowledge distillation. In Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval, 546--555

    2022

  15. [23]

    Ye, Y.; Xia, L.; and Huang, C. 2023. Graph masked autoencoder for sequential recommendation. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 321--330

    2023

  16. [24]

    Zhang, S.; Liu, Y.; Sun, Y.; and Shah, N. 2021. Graph-less neural networks: Teaching old mlps new tricks via distillation. arXiv preprint arXiv:2110.08727

    2021 arXiv

  17. [25]

    Zhang, Y.; Liu, Y.; Xu, Y.; Xiong, H.; Lei, C.; He, W.; Cui, L.; and Miao, C. 2022. Enhancing sequential recommendation with graph contrastive learning. arXiv preprint arXiv:2205.14837

    2022 arXiv

  18. [26]

    Zhang, Y.; Wu, Y.; Le, R.; Zhu, Y.; Zhuang, F.; Han, R.; Li, X.; Lin, W.; An, Z.; and Xu, Y. 2023. Modeling dual period-varying preferences for takeaway recommendation. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 5628--5638

    2023

  19. [27]

    Zhu, Q.; Chen, X.; Wu, P.; Liu, J.; and Zhao, D. 2021. Combining curriculum learning and knowledge distillation for dialogue generation. In Findings of the Association for Computational Linguistics: EMNLP 2021, 1284--1295

    2021

  20. [28]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  21. [29]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.