Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Gwhere, an LLM-based generative recommender that assigns each place a learned semantic identifier and generates the user's next POI, beats Amap's production cascade-ranking system offline and online (P-CTR +5.83%, U-CTR +6.20%).

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

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

T0 review · deepseek-v4-flash

2026-08-02 06:47 UTC pith:DADK2C7U

load-bearing objection Potentially solid industrial gen-rec paper, but the ALS tokenizer may leak test labels; check the split before trusting offline Acc@1. the 3 major comments →

arxiv 2607.26073 v1 pith:DADK2C7U submitted 2026-07-13 cs.IR

Guess Where You Go: Generative Next Point-of-Interest Recommendation in Amap

Penglong Zhai , Bowen Zheng , Jie Li , Yifang Yuan , Yue Liu , Sicong Wang , Mingyang Yin , Tingting Hu
show 3 more authors
Shuaijun Guo Fanyi Di Xin Li
This is my paper
classification cs.IR
keywords generative retrievalnext POI recommendationsemantic identifierscontrastive tokenizationresidual quantizationlarge language modelspreference optimizationonline A/B testing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that next-POI recommendation can be cast as a language-generation task: assign each place a compact semantic identifier (SID) distilled from text, image, geography, and collaborative signals, then fine-tune an LLM to generate the SID of the user's next destination. The authors claim that this generative pipeline outperforms both prior next-POI baselines and Amap's production cascade-ranking system offline, and that a one-month online A/B test shows relative gains of 5.83% in page CTR and 6.20% in user CTR. The paper also introduces Exposure-Aware Kahneman-Tversky Optimization (EAKTO), a preference-alignment objective that treats clicked POIs as desirable and exposed-but-unclicked as noisy negatives, and reports that removing it halves the online gains and worsens negative-feedback rate. If true, the result suggests that industrial-scale location-based recommendation can move from multi-stage retriever-ranker pipelines to a single generative model.

Core claim

Gwhere claims that next-POI recommendation can be cast as generative retrieval: each POI gets a compact Semantic ID (SID) learned by a contrastive residual-quantization tokenizer over textual, visual, spatial, and collaborative embeddings; an LLM is post-trained with continued pretraining, SFT, and Exposure-Aware Kahneman-Tversky Optimization to autoregressively generate the SID of the user's next visit. On NYC, TKY, Gowalla-CA, Gwhere-0.5B surpasses prior generative SOTA on Acc@1; on Amap's industrial set, all variants beat production cascade ranking. The deployed 0.5B model reports one-month online lifts of 5.83% P-CTR and 6.20% U-CTR over the cascade baseline.

What carries the argument

The Semantic ID (SID) produced by a contrastive residual-quantization tokenizer. Each modality embedding (text, image, spatial, collaborative) is projected and attention-fused, then passed through L layers of Gumbel-softmax residual quantization; an NT-Xent contrastive loss replaces reconstruction loss to keep codes discriminative and avoid code collapse. The SID converts next-POI prediction into autoregressive generation of discrete tokens. The second load-bearing piece is EAKTO, a KTO-style loss with positive reward anchoring and conditional negative updates, which uses exposure logs (clicks as desirable, exposed-unclicked as negatives) without requiring paired preference data.

Load-bearing premise

The collaborative embedding that defines each POI's semantic ID is fit by ALS on the user–POI interaction matrix, and the paper never states that this matrix excludes the held-out target interaction of the temporal test split; if the label leaks into the SID, the reported gains are inflated.

What would settle it

Recompute Gwhere's offline Acc@1 after fitting the ALS collaborative embeddings on the training-interval interactions only (80% of the temporal split), holding everything else constant; if Acc@1 drops materially, the original results are partly leakage. Also audit the released pipeline for whether test-split target check-ins ever enter the user–POI matrix.

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

If this is right

  • Removing the semantic-ID tokenizer is the single largest ablation (-16.01% Acc@1 on AMAP), so the contrastive SID, not the LLM alone, carries much of the predictive signal.
  • EAKTO outperforms DPO and GRPO for preference alignment, and the online ablation shows its removal roughly halves CTR gains and weakens negative-feedback suppression.
  • Scaling the base LLM from 0.5B to 7B monotonically improves offline Acc@1, but the deployed model is limited to 0.5B by the ~30ms p99 latency budget at 50 QPS on two H20 GPUs.
  • If the claimed gains hold generally, the multi-stage retriever-coarse-fine-cognitive ranker can be replaced by a single generative model in high-concurrency LBS recommendations.

Where Pith is reading between the lines

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

  • The paper does not state whether the ALS collaborative matrix used in SID construction is restricted to training-time check-ins; if the held-out target interaction participates in fitting the SID, the tokenizer leaks the label into the code the LLM must generate, inflating both offline Acc@1 and online lift. A temporal split on the ALS input is the natural test.
  • The online baseline is Amap's production cascade system, which optimizes engagement and may not match the generative model's training objective; the +5.83%/+6.20% lift could partly reflect measurement or optimization-target differences rather than the generative paradigm per se.
  • The nearly linear Acc@1 gain with continued-pretraining corpus size implies the framework is currently data-bound; scaling the anonymized trajectory corpus, not model size, is the cheapest path to better predictions.
  • The same SID-plus-LLM recipe should transfer to other spatial entity generation tasks (trip planning, route recommendation, multi-destination itineraries), where the contrastive spatio-temporal tokenizer would replace discrete IDs.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes Gwhere, an industrial framework for generative next-POI recommendation. It first learns compact semantic identifiers (SIDs) through a contrastive residual-quantization tokenizer that fuses textual, visual, spatial, and collaborative (ALS) signals, then adapts an LLM (Qwen2.5) to POI prediction via continued pretraining, SFT, and a proposed Exposure-Aware Kahneman-Tversky Optimization (EAKTO) objective. Experiments on Foursquare-NYC/TKY, Gowalla-CA, and a proprietary Amap dataset compare Gwhere against classical and LLM-based baselines, report offline Acc@1 gains (e.g., 0.4027 vs. 0.3618 on NYC), and present a one-month online A/B test on Amap's homepage showing +5.83% P-CTR and +6.20% U-CTR over the production cascade-ranking baseline. The implementation is claimed to be publicly available.

Significance. If the claims hold, the paper makes a strong contribution: it demonstrates that LLM-based generative retrieval with learned semantic IDs can replace cascade ranking in a large-scale production next-POI system, while also proposing a tokenizer that explicitly targets discriminability rather than reconstruction. The public-code commitment and the detailed industrial deployment description are valuable. However, the central evidence rests on (i) a data-splitting assumption about the ALS collaborative embeddings that is not stated, and (ii) point estimates without uncertainty quantification. These issues must be resolved before the effectiveness claims can be accepted.

major comments (3)
  1. [§3.2.1, §4.1.1, §4.1.2] The collaborative embedding z_cf is described as obtained via ALS on the user-POI interaction matrix, but the paper never states that this matrix is restricted to the training portion of the 80/10/10 temporal split. Since §4.1.1 holds out each user's last visit as ground truth, if the ALS matrix includes that held-out interaction, the target POI's SID is computed from the very label the model is asked to predict, and the contrastive tokenizer propagates this leakage into the SIDs. This would directly inflate Acc@1 in Tables 3–5 and the claimed gains over GNPR-SID. Please state explicitly that the interaction matrix is train-only, and ideally rerun the public-dataset evaluations with a strictly train-only ALS fit.
  2. [Table 4, §4.2] The headline offline comparison against baselines depends on Acc@1 point estimates. Baseline numbers are 'borrowed from [13,26]' with no variance, and Gwhere numbers are also reported as single points without confidence intervals or significance tests. The margins over GNPR-SID (e.g., 0.0409 on NYC) may be real, but without uncertainty or a description of how the borrowed baselines were computed on the same split, the SOTA claim is not statistically grounded. Report standard deviations or significance tests, or at least state whether the baseline numbers were recomputed under identical preprocessing.
  3. [§3.4, §4.4.3, Table 6] EAKTO is trained to maximize the likelihood of clicked POIs over exposed-but-unclicked POIs, and the online A/B's primary metrics are CTR. Part of the reported +5.83% P-CTR / +6.20% U-CTR lift is therefore a direct test of the training objective, which is a mild circularity. The claim would be strengthened by reporting significance intervals for the A/B metrics and by emphasizing downstream metrics less aligned with the objective (e.g., retention, negative-feedback rate). As written, the online results are self-reported point estimates without confidence intervals.
minor comments (5)
  1. [Table 5] The right panel of Table 5 does not state which Gwhere variant is used (the left panel says Gwhere-3B, but the right panel's SFT+EAKTO value 0.2820 appears to match Gwhere-0.5B in Table 3). Please clarify the model size for the preference-alignment comparison.
  2. [§4.1.2] The codebook sizes are described as 'tens of codewords each (e.g., 48 for public datasets, larger for AMAP)' — please give the exact K and L values used for each dataset, since token collision rate and SID length directly affect the generative retrieval results.
  3. [Figure 3] The axes of Figure 3 are not described in the caption; please label them (e.g., input sequence length vs. Acc@1, training tokens vs. Acc@1) so the saturation claim can be checked.
  4. [Abstract / §1] The public repository link points to 'SimCIT', which is not mentioned in the paper body. Please either align the repository name with Gwhere or explain the relationship.
  5. [§5] The conclusion mentions 'prefill-decoding decoupling and multi-token prediction' as engineering optimizations, but these are not described in the experimental or deployment sections. Please either add details or remove the unsupported claim.

Circularity Check

0 steps flagged

No significant circularity: Gwhere's training/evaluation chain is not defined in terms of its own outputs, and the cited prior work is not load-bearing.

full rationale

The derivation chain is: (1) learn POI SIDs via contrastive residual quantization over textual, visual, spatial, and collaborative embeddings; (2) continuously pretrain an LLM on SID-location-description and spatio-temporal behavior corpora; (3) SFT on next-POI prompt/response pairs; (4) EAKTO on exposure logs; (5) evaluate offline on temporally split data and online via prospective A/B testing. None of these stages defines the target quantity in terms of itself. EAKTO does optimize click/non-click exposure signals, and online P-CTR/U-CTR measure clicks, but the online A/B test is held-out traffic and the ablation is compared against DPO/GRPO, so the reported CTR lift is an empirical result rather than a fitted parameter renamed as a prediction. The ALS-based collaborative embedding is described without an explicit statement that the user-POI matrix is restricted to the training split; if the matrix included held-out last visits, this would be label leakage and could inflate Acc@1, but that is a data-splitting concern, not a demonstrated circular step, and the paper's split description ('the last visited POI is held out as the ground truth') gives no direct evidence that ALS sees test interactions. Self-citations such as OneRec and LC-Rec appear only in related work and are not used to justify the paper's central claims; there is no imported uniqueness theorem or ansatz-by-citation chain. The framework is therefore self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central claim is an empirical system result, so most load-bearing choices are hyperparameters and data-processing assumptions rather than mathematical axioms. The largest unpaid assumption is that the collaborative signals used to build SIDs are temporally clean; the paper does not document this.

free parameters (4)
  • SID codebook sizes (K per layer) and number of codebooks L = K=48 per layer for public datasets; larger, unspecified for AMAP; L=3
    Chosen by hand; directly determines granularity and collision behavior of the semantic IDs that the whole generative pipeline is built on (§3.2.2, §4.1.2).
  • NT-Xent temperature τ = 0.1
    Fixed hyperparameter in the contrastive tokenizer loss (§3.2.2); affects separability of SIDs.
  • EAKTO hyperparameters λ_D, λ_U, β, α = 1.0, 1.0, 0.1, 0.3
    Chosen by hand for the preference-alignment loss in Eq. (1); no sensitivity analysis or selection procedure reported (§3.4).
  • Latent dimension and Gumbel-softmax annealing schedule = 96-dim latent; α annealed but schedule not specified
    Architectural choices that affect quantization and training stability (§4.1.2).
axioms (5)
  • standard math Gumbel-softmax relaxation is a valid differentiable proxy for discrete code assignment
    Used in §3.2.2 to backprop through residual quantization; standard but approximate, and annealing schedule is not specified.
  • domain assumption Click = desirable, exposed-unclicked = undesirable is a valid preference signal
    Core to EAKTO in §3.4; the paper itself notes noisy negatives (inattention, latency) and adds a heuristic mask rather than modeling them, so the assumption is partially acknowledged.
  • domain assumption ALS collaborative embeddings are computed without the held-out test interaction
    Not stated anywhere (§3.2.1 vs §4.1.1); if false, Acc@1 and CTR gains are inflated by label leakage.
  • domain assumption Six-character Geohash captures useful spatial proximity
    Used to align SIDs to location in continued pretraining (§3.3.1); coarse Geohash may not capture fine-grained POI proximity.
  • domain assumption Narrativized check-in corpora transfer to real next-POI prediction
    Continued pretraining uses synthetic natural-language sequences of desensitized trajectories (§3.3.1); the paper does not measure transfer separately from SFT.

pith-pipeline@v1.3.0-alltime-deepseek · 15059 in / 14851 out tokens · 134471 ms · 2026-08-02T06:47:31.469312+00:00 · methodology

0 comments
read the original abstract

Generative retrieval enables recommender systems to retrieve items by generating compact item identifiers, but scaling it to industrial scenarios remains challenging due to redundant or colliding token assignments and insufficient integration of heterogeneous item signals. These challenges are particularly critical for next Point-of-Interest (POI) recommendation, where models must represent structured spatial entities, capture sequential mobility patterns, and produce predictions consistent with real user behavior. We propose Gwhere, an end-to-end industrial framework that integrates semantic identifier (SID) generation with LLM-based generative next POI recommendation. Gwhere first learns discriminative POI SIDs through a contrastive residual-quantization tokenizer that aligns textual, visual, spatial, and collaborative signals. Based on these SIDs, Gwhere adapts LLMs to mobility scenarios via continued pretraining on enriched spatio-temporal corpora, supervised fine-tuning, and Exposure-Aware Kahneman-Tversky Optimization (EAKTO), a reinforcement learning objective for behavioral preference alignment. Experiments on public datasets and Amap's large-scale industrial dataset demonstrate the effectiveness of Gwhere. The system has been deployed in Amap's homepage service under high-concurrency and low-latency constraints. Long-term online A/B tests show improvements of 5.83% in P-CTR and 6.20% in U-CTR over the production baseline. The implementation is publicly available at https://github.com/alibaba/SimCIT.

Figures

Figures reproduced from arXiv: 2607.26073 by Bowen Zheng, Fanyi Di, Jie Li, Mingyang Yin, Penglong Zhai, Shuaijun Guo, Sicong Wang, Tingting Hu, Xin Li, Yifang Yuan, Yue Liu.

Figure 1
Figure 1. Figure 1: Offline Training and Online Serving Procedure of proposed framework Gwhere. Offline Training:(1) POIs are tokenized [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Impact on the performance of sequence length [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of Model Inference p99 Latency (ms) [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

37 extracted references · 6 linked inside Pith

  1. [1]

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. 2020. A simple framework for contrastive learning of visual representations. InInterna- tional Conference on Machine Learning. Guess Where You Go: Generative Next Point-of-Interest Recommendation in Amap RecSys ’26, September 28–October 2, 2026, Minneapolis, Minnesota, USA

    2020

  2. [2]

    Myers, and Jure Leskovec

    Eunjoon Cho, Seth A. Myers, and Jure Leskovec. 2011. Friendship and mobility: user movement in location-based social networks. InProceedings of the 17th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining(San Diego, California, USA)(KDD ’11). Association for Computing Machinery, New York, NY, USA, 1082–1090. doi:10.1145/2020408.2020579

    arXiv 2011

  3. [3]

    Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa, René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. 2024. A review of modern recommender systems using generative models (gen-recsys). InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 6448–6458

    2024

  4. [4]

    Jiaxin Deng, Shiyao Wang, Kuo Cai, Lejian Ren, Qigen Hu, Weifeng Ding, Qiang Luo, and Guorui Zhou. 2025. OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment. arXiv:2502.18965 [cs.IR] https://arxiv.org/abs/2502.18965

    Pith/arXiv arXiv 2025

  5. [5]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American chapter of the Association for Computational Linguistics: Human Language Technologies. 4171–4186

    2019

  6. [6]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. 2021. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. arXiv:2010.11929 [cs.CV] https://arxiv.org/abs/2010.11929

    Pith/arXiv arXiv 2021

  7. [7]

    Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. 2024. Model alignment as prospect theoretic optimization. InProceedings of the 41st International Conference on Machine Learning(Vienna, Austria)(ICML’24). JMLR.org, Article 504, 18 pages

    2024

  8. [8]

    Chao Feng, Wuchao Li, Defu Lian, Zheng Liu, and Enhong Chen. 2022. Recom- mender forest for efficient retrieval. InAdvances in Neural Information Processing Systems

    2022

  9. [9]

    Shanshan Feng, Xutao Li, Yifeng Zeng, Gao Cong, Yeow Meng Chee, and Quan Yuan. 2015. Personalized ranking metric embedding for next new POI recommen- dation. InProceedings of the 24th International Conference on Artificial Intelligence (Buenos Aires, Argentina)(IJCAI’15). AAAI Press, 2069–2075

    2015

  10. [10]

    Shanshan Feng, Feiyu Meng, Lisi Chen, Shuo Shang, and Yew Soon Ong. 2024. ROTAN: A Rotation-based Temporal Attention Network for Time-Specific Next POI Recommendation. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining(Barcelona, Spain)(KDD ’24). Association for Computing Machinery, New York, NY, USA, 759–770. doi:10.11...

    doi:10.1145/3637528 2024

  11. [11]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2022. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). InProceedings of the 16th ACM Conference on Recommender Systems(Seattle, WA, USA)(RecSys ’22). Association for Computing Machinery, New York, NY, USA, 299–315. doi:10.11...

    arXiv 2022

  12. [12]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. 2017. Inductive representation learning on large graphs. InAdvances in Neural Information Processing Systems. 1025–1035

    2017

  13. [13]

    Peibo Li, Maarten de Rijke, Hao Xue, Shuang Ao, Yang Song, and Flora D. Salim. 2024. Large Language Models for Next Point-of-Interest Recommen- dation. InProceedings of the 47th International ACM SIGIR Conference on Re- search and Development in Information Retrieval(Washington DC, USA)(SIGIR ’24). Association for Computing Machinery, New York, NY, USA, 1...

    arXiv 2024

  14. [14]

    Yan Luo, Haoyi Duan, Ye Liu, and Fu-Lai Chung. 2023. Timestamps as Prompts for Geography-Aware Location Recommendation. InProceedings of the 32nd ACM International Conference on Information and Knowledge Management(Birming- ham, United Kingdom)(CIKM ’23). Association for Computing Machinery, New York, NY, USA, 1697–1706. doi:10.1145/3583780.3615083

    arXiv 2023

  15. [15]

    Yingtao Luo, Qiang Liu, and Zhaocheng Liu. 2021. STAN: Spatio-Temporal Attention Network for Next Location Recommendation. InProceedings of the Web Conference 2021(Ljubljana, Slovenia)(WWW ’21). Association for Computing Machinery, New York, NY, USA, 2177–2185. doi:10.1145/3442381.3449998

    arXiv 2021

  16. [16]

    Cheonbok Park, Chunggi Lee, Hyojin Bahng, Yunwon Tae, Seungmin Jin, Ki- hwan Kim, Sungahn Ko, and Jaegul Choo. 2020. ST-GRAT: A Novel Spatio- temporal Graph Attention Networks for Accurately Forecasting Dynamically Changing Road Speed. InProceedings of the 29th ACM International Conference on Information & Knowledge Management (CIKM ’20). ACM, 1215–12...

    arXiv 2020

  17. [17]

    Haohao Qu, Wenqi Fan, Zihuai Zhao, and Qing Li. 2024. TokenRec: Learning to Tokenize ID for LLM-based Generative Recommendation

    2024

  18. [18]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li,...

    Pith/arXiv arXiv 2025

  19. [19]

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. InThirty-seventh Conference on Neural Infor- mation Processing Systems. https://openreview.net/forum?id=HPuSIXJaa9

    2023

  20. [20]

    Tran, Jonah Samost, Maciej Kula, Ed H

    Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Keshavan, Trung Vu, Lukasz Heidt, Lichan Hong, Yi Tay, Vinh Q. Tran, Jonah Samost, Maciej Kula, Ed H. Chi, and Maheswaran Sathiamoorthy. 2023. Recommender systems with generative retrieval. InAdvances in Neural Information Processing Systems

    2023

  21. [21]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. 2024. DeepSeek- Math: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300 [cs.CL] https://arxiv.org/abs/2402.03300

    Pith/arXiv arXiv 2024

  22. [22]

    Anima Singh, Trung Vu, Nikhil Mehta, Raghunandan Keshavan, Maheswaran Sathiamoorthy, Yilin Zheng, Lichan Hong, Lukasz Heldt, Li Wei, Devansh Tandon, Ed Chi, and Xinyang Yi. 2024. Better Generalization with Semantic IDs: A Case Study in Ranking for Recommendations. InProceedings of the 18th ACM Conference on Recommender Systems. 1039–1044

    2024

  23. [23]

    Hao Sun, Changjie Yang, Liwei Deng, Fan Zhou, Feiteng Huang, and Kai Zheng

  24. [24]

    Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W

    Yi Tay, Vinh Q. Tran, Mostafa Dehghani, Jianmo Ni, Dara Bahri, Harsh Mehta, Zhen Qin, Kai Hui, Zhe Zhao, Jai Gupta, Tal Schuster, William W. Cohen, and Donald Metzler. 2022. Transformer memory as a differentiable search index. In Advances in Neural Information Processing Systems

    2022

  25. [25]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InAdvances in Neural Information Processing Systems, Vol. 30

    2017

  26. [26]

    Dongsheng Wang, Yuxi Huang, Shen Gao, Yifan Wang, Chengrui Huang, and Shuo Shang. 2025. Generative Next POI Recommendation with Semantic ID. arXiv:2506.01375 [cs.IR] https://arxiv.org/abs/2506.01375

    Pith/arXiv arXiv 2025

  27. [27]

    Wenjie Wang, Honghui Bao, Xinyu Lin, Jizhi Zhang, Yongqi Li, Fuli Feng, See- Kiong Ng, and Tat-Seng Chua. 2024. Learnable Item Tokenization for Generative Recommendation. InProceedings of the 33rd ACM International Conference on Information and Knowledge Management(Boise, ID, USA)(CIKM ’24). Association for Computing Machinery, New York, NY, USA, 2400–2409

    2024

  28. [28]

    Ye Wang, Jiahao Xun, Minjie Hong, Jieming Zhu, Tao Jin, Wang Lin, Haoyuan Li, Linjun Li, Yan Xia, Zhou Zhao, and Zhenhua Dong. 2024. EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 3245–3254

    2024

  29. [29]

    Yuxia Wu, Ke Li, Guoshuai Zhao, and Xueming Qian. 2022. Personalized Long- and Short-term Preference Learning for Next POI Recommendation.IEEE Transactions on Knowledge & Data Engineering34, 04 (April 2022), 1944–1957. doi:10.1109/TKDE.2020.3002531

    arXiv 2022

  30. [30]

    Xiaodong Yan, Tengwei Song, Yifeng Jiao, Jianshan He, Jiaotuan Wang, Ruopeng Li, and Wei Chu. 2023. Spatio-Temporal Hypergraph Learning for Next POI Recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval(Taipei, Taiwan)(SIGIR ’23). Association for Computing Machinery, New York, NY,...

    arXiv 2023

  31. [31]

    Zheng, and Zhiyong Yu

    Dingqi Yang, Daqing Zhang, Vincent W. Zheng, and Zhiyong Yu. 2015. Modeling User Activity Preference by Leveraging User Spatial Temporal Characteristics in LBSNs.IEEE Transactions on Systems, Man, and Cybernetics: Systems45, 1 (2015), 129–142

    2015

  32. [32]

    Song Yang, Jiamou Liu, and Kaiqi Zhao. 2022. GETNext: Trajectory Flow Map Enhanced Transformer for Next POI Recommendation. InProceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval(Madrid, Spain)(SIGIR ’22). Association for Computing Machinery, New York, NY, USA, 1144–1153. doi:10.1145/3477495.3531983

    arXiv 2022

  33. [33]

    Bowen Zheng, Yupeng Hou, Hongyu Lu, Yu Chen, Wayne Xin Zhao, Ming Chen, and Ji-Rong Wen. 2024. Adapting Large Language Models by Integrating Collaborative Semantics for Recommendation.arXiv preprint arXiv:2311.09049 (2024)

    Pith/arXiv arXiv 2024

  34. [34]

    Guorui Zhou, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Qiang Luo, Qianqian Wang, Qigen Hu, Rui Huang, Shiyao Wang, Weifeng Ding, Wuchao Li, Xinchen Luo, Xingmei Wang, Zexuan Cheng, Zixing Zhang, Bin Zhang, Boxuan Wang, Chaoyi Ma, Chengru Song, Chenhui Wang, Di Wang, Dongxue Meng, Fan Yang, Fangyu Zhang, Feng Jiang, Fuxing Zhang, Gang Wang, Guowang ...

    arXiv 2025

  35. [35]

    Jieming Zhu, Mengqun Jin, Qijiong Liu, Zexuan Qiu, Zhenhua Dong, and Xiu Li

  36. [2021]

    InProceedings of the 30th ACM International Conference on Information & Knowledge Management(Virtual Event, Queensland, Australia)(CIKM ’21)

    PeriodicMove: Shift-aware Human Mobility Recovery with Graph Neural Network. InProceedings of the 30th ACM International Conference on Information & Knowledge Management(Virtual Event, Queensland, Australia)(CIKM ’21). Association for Computing Machinery, New York, NY, USA, 1734–1743. doi:10. 1145/3459637.3482284

    arXiv

  37. [2024]

    InProceedings of the 18th ACM Conference on Recommender Systems(Bari, Italy)(RecSys ’24)

    CoST: Contrastive Quantization based Semantic Tokenization for Genera- tive Recommendation. InProceedings of the 18th ACM Conference on Recommender Systems(Bari, Italy)(RecSys ’24). Association for Computing Machinery, New York, NY, USA, 969–974. doi:10.1145/3640457.3688178

    arXiv