arxiv: 2604.23156 · v1 · submitted 2026-04-25 · 💻 cs.IR

Recognition: unknown

Birds of a Feather Cluster Nearby: a Proximity-Aware Geo-Codebook for Local Service Recommendation

Tian He , Chen Yang , Jiawei Zhang , Lin Guo , Wei Lin , Zhuqing Jiang

Authors on Pith no claims yet

Pith reviewed 2026-05-08 07:29 UTC · model grok-4.3

classification 💻 cs.IR

keywords proximity-aware geo-codebooklocal service recommendationgenerative recommendationgeo-centroid coordinate systemgeo-rotary position encodingsemantic ID tokenizationgeographic feasibility

0 comments

The pith

Pro-GEO builds a geo-centroid coordinate system and geo-rotary encoding to jointly capture semantic relevance and geographic proximity in local service recommendations.

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

The paper introduces Pro-GEO to fix a gap in generative recommendation systems for local services. Semantic ID tokenization often produces relevant but unreachable suggestions because it ignores location constraints. Pro-GEO creates a local geo-centroid coordinate system inside clusters and uses geo-rotary position encoding to treat proximity as rotations in embedding space. This lets semantic and spatial signals combine without geography becoming a secondary add-on. Tests on a large industrial dataset show clear gains in clustering distance and hit rate.

Core claim

Pro-GEO establishes a geo-centroid local coordinate system to capture intra-cluster spatial relationships and a geo-rotary position encoding mechanism that models geographic proximity as orthogonal rotational transformations in the high-dimensional embedding. This design enables semantic and spatial signals to be jointly modeled in a balanced manner, without reducing geographic information to a weak auxiliary feature.

What carries the argument

The geo-rotary position encoding mechanism that represents geographic proximity as orthogonal rotational transformations inside high-dimensional embeddings while using a geo-centroid local coordinate system for intra-cluster relations.

If this is right

Semantic ID tokenization now respects strict geographic feasibility for local services instead of producing unreachable items.
Average geographic clustering distance drops by 45.60% on large-scale industrial data.
Hit@50 improves by 1.87% over prior state-of-the-art methods.
Semantic and spatial signals remain balanced without geography treated as a weak auxiliary input.

Where Pith is reading between the lines

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

The rotational view of proximity could be tested in other spatial embedding tasks such as map-based search or location-aware advertising.
If the encoding preserves performance across cities with different density patterns, it would support broader deployment beyond the original industrial setting.
The same coordinate-plus-rotation idea might simplify fusion of other real-world constraints like time or cost into embedding spaces.

Load-bearing premise

That the geo-rotary position encoding can jointly model semantic and spatial signals in a balanced manner without geographic information being reduced to a weak auxiliary feature, and that this design generalizes from the industrial dataset to other local service scenarios.

What would settle it

Running the same model on a separate local-service dataset and finding no reduction in average geographic clustering distance or no gain in Hit@50 relative to standard semantic codebooks.

Figures

Figures reproduced from arXiv: 2604.23156 by Chen Yang, Jiawei Zhang, Lin Guo, Tian He, Wei Lin, Zhuqing Jiang.

**Figure 1.** Figure 1: Illustration of local lifestyle recommendation. view at source ↗

**Figure 2.** Figure 2: The overview of the ProGEO. It includes two standard codebook layers and a geo-codebook layer. The standard view at source ↗

**Figure 3.** Figure 3: Comparison between global and local Coordinate view at source ↗

**Figure 4.** Figure 4: Comparison of global and local geographic repre view at source ↗

**Figure 5.** Figure 5: Comparison of Geo-RoPE integration strategies at view at source ↗

**Figure 6.** Figure 6: Comparison of geographic information enhance view at source ↗

**Figure 8.** Figure 8: Visualization of codebook geographical clustering view at source ↗

**Figure 9.** Figure 9: Visualization of recommended POI distributions view at source ↗

**Figure 7.** Figure 7: presents the sensitivity analysis of the rotation scale parameters (𝛼, 𝛽) in the Geo-RoPE process across six evaluation metrics. As (𝛼, 𝛽) increase, all distance-based metrics (average distance, p90 distance, and p95 distance) exhibit a pronounced decline, suggesting that larger rotation scales facilitate more compact and discriminative spatial clustering. Specifically, the average distance decreases fro… view at source ↗

**Figure 10.** Figure 10: Additional case studies on the spatial distribution of POI recommendation results. E view at source ↗

read the original abstract

Generative recommendation systems are increasingly adopted in local service platforms, where semantic relevance alone is insufficient without strict geographic feasibility. A key technical challenge lies in semantic ID (SID) tokenization, which directly impacts the recommendation performance. However, existing semantic codebooks neglect geographic constraints, often resulting in recommendations that are semantically relevant yet geographically unreachable. To address this limitation, we propose Pro-GEO, a Proximity-aware GEO-codebook. Pro-GEO establishes a geo-centroid local coordinate system to capture intra-cluster spatial relationships and a geo-rotary position encoding mechanism that models geographic proximity as orthogonal rotational transformations in the high-dimensional embedding. This design enables semantic and spatial signals to be jointly modeled in a balanced manner, without reducing geographic information to a weak auxiliary feature. Extensive experiments conducted on a large-scale industrial dataset reveal that Pro-GEO significantly outperforms state-of-the-art methods. In particular, Pro-GEO reduces the average geographic clustering distance by 45.60% and achieves a 1.87% improvement in Hit@50, highlighting its effectiveness for real-world local service recommendation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Pro-GEO adds a local geo-centroid system and rotary encodings to semantic codebooks so proximity is baked into embeddings rather than treated as an afterthought, with concrete gains on industrial data but sparse experimental details.

read the letter

The main thing to know is that this paper gives semantic ID tokenization a geographic upgrade for local service recommenders. It builds a geo-centroid local coordinate system to handle distances inside clusters and uses geo-rotary position encoding to turn proximity into rotational transformations in the embedding space. That setup lets semantic and spatial signals sit together without location getting pushed to the side, which directly tackles the problem of recommendations that look good but are too far away to use.

Referee Report

2 major / 2 minor

Summary. The paper claims that semantic ID tokenization in generative recommendation systems for local services often ignores geographic constraints, leading to semantically relevant but unreachable recommendations. To address this, it proposes Pro-GEO, which introduces a geo-centroid local coordinate system for capturing intra-cluster spatial relationships and a geo-rotary position encoding mechanism that represents geographic proximity via orthogonal rotational transformations in high-dimensional embeddings. This design jointly models semantic and spatial signals without treating geography as a weak auxiliary feature. On a large-scale industrial dataset, Pro-GEO reduces average geographic clustering distance by 45.60% and improves Hit@50 by 1.87% over state-of-the-art methods.

Significance. If the experimental claims hold under rigorous validation, this work offers a meaningful advance for local service platforms by balancing semantic relevance with geographic feasibility, a key requirement in domains such as food delivery or local search. The extension of rotary embeddings to encode proximity as rotations is a technically interesting idea that could inspire further work on spatially aware tokenization. The provision of concrete performance deltas on an industrial dataset is a positive aspect, as is the focus on avoiding dimensional collapse between signals.

major comments (2)

[Experiments] Experiments section: The reported gains (45.60% reduction in geographic clustering distance and 1.87% Hit@50 improvement) are central to the paper's contribution, yet the manuscript provides no details on the specific baselines compared, the exact implementation of the geo-rotary position encoding (e.g., how the orthogonal transformations are parameterized or integrated with SID embeddings), or any statistical significance tests and confidence intervals for the improvements. This absence makes it impossible to assess whether the results support the claims or rule out confounds such as dataset-specific tuning.
[Method] Method section (geo-rotary position encoding): The description states that geographic proximity is modeled as orthogonal rotational transformations, but it is unclear how this mechanism ensures balanced joint modeling of semantic and spatial signals without one dominating (e.g., via explicit loss terms, hyperparameter schedules, or ablation studies isolating the rotary component). If the encoding reduces to a weak auxiliary signal in practice, the core design claim would not hold.

minor comments (2)

[Abstract] Abstract and introduction: Acronyms such as SID should be expanded on first use for clarity.
[Figures/Tables] Ensure that all figures and tables include clear captions explaining axes, metrics, and any error bars or statistical annotations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback, which highlights important areas for improving the clarity and rigor of our manuscript. We address each major comment in detail below and will revise the paper accordingly to provide the requested information and analyses.

read point-by-point responses

Referee: [Experiments] Experiments section: The reported gains (45.60% reduction in geographic clustering distance and 1.87% Hit@50 improvement) are central to the paper's contribution, yet the manuscript provides no details on the specific baselines compared, the exact implementation of the geo-rotary position encoding (e.g., how the orthogonal transformations are parameterized or integrated with SID embeddings), or any statistical significance tests and confidence intervals for the improvements. This absence makes it impossible to assess whether the results support the claims or rule out confounds such as dataset-specific tuning.

Authors: We agree that additional experimental details are essential for validating the reported improvements. In the revised manuscript, we will expand the Experiments section to include: (1) a complete description of all baselines, including their implementations, hyperparameters, and references; (2) the full mathematical formulation of the geo-rotary position encoding, specifying how orthogonal transformations are parameterized (e.g., via rotation matrices derived from geographic coordinates) and integrated with SID embeddings; and (3) statistical significance tests (e.g., paired t-tests) with confidence intervals for the key metrics. These additions will allow readers to rigorously evaluate the results and rule out potential confounds. revision: yes
Referee: [Method] Method section (geo-rotary position encoding): The description states that geographic proximity is modeled as orthogonal rotational transformations, but it is unclear how this mechanism ensures balanced joint modeling of semantic and spatial signals without one dominating (e.g., via explicit loss terms, hyperparameter schedules, or ablation studies isolating the rotary component). If the encoding reduces to a weak auxiliary signal in practice, the core design claim would not hold.

Authors: The geo-rotary position encoding models proximity through orthogonal rotations in the shared embedding space, which by construction preserves semantic directions while incorporating spatial information without dimensional collapse. However, we acknowledge that the current Method section does not sufficiently detail the balancing process or provide supporting analyses. In the revision, we will clarify the integration mechanism, specify any hyperparameter schedules used for balancing, and add ablation studies that isolate the geo-rotary component (e.g., comparing variants with and without the encoding). This will demonstrate that spatial signals are not reduced to a weak auxiliary feature but are jointly optimized with semantics. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper proposes Pro-GEO as a new codebook architecture that combines a geo-centroid local coordinate system with geo-rotary position encoding to jointly embed semantic IDs and geographic proximity. These components are introduced as explicit design choices rather than derived from prior fitted quantities or self-referential definitions. The central performance claims (45.60% reduction in average geographic clustering distance and 1.87% Hit@50 gain) rest on empirical evaluation against baselines on an industrial dataset, with no equations or steps shown that reduce by construction to the inputs, no load-bearing self-citations, and no renaming of known results as novel derivations. The derivation chain is therefore self-contained and non-circular.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The approach relies on standard embedding assumptions and introduces two new mechanisms without independent evidence beyond the reported experiments.

axioms (1)

domain assumption High-dimensional embeddings can represent semantic and spatial signals jointly when using rotational transformations for proximity.
Invoked in the description of the geo-rotary position encoding mechanism.

invented entities (2)

geo-centroid local coordinate system no independent evidence
purpose: Capture intra-cluster spatial relationships
Newly proposed component of Pro-GEO.
geo-rotary position encoding no independent evidence
purpose: Model geographic proximity as orthogonal rotational transformations
Newly proposed mechanism of Pro-GEO.

pith-pipeline@v0.9.0 · 5498 in / 1379 out tokens · 46397 ms · 2026-05-08T07:29:01.802834+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

35 extracted references · 18 canonical work pages · 4 internal anchors

[1]

Ben Chen, Xian Guo, Siyuan Wang, Zihan Liang, Yue Lv, Yufei Ma, Xinlong Xiao, Bowen Xue, Xuxin Zhang, Ying Yang, et al . 2025. Onesearch: A preliminary exploration of the unified end-to-end generative framework for e-commerce search.arXiv preprint arXiv:2509.03236(2025)

work page arXiv 2025
[2]

Jiahui Chen, Xiaoze Jiang, Zhibo Wang, Quanzhi Zhu, Junyao Zhao, Feng Hu, Kang Pan, Ao Xie, Maohua Pei, Zhiheng Qin, et al. 2025. UniSearch: Rethink- ing Search System with a Unified Generative Architecture.arXiv preprint arXiv:2509.06887(2025)

work page arXiv 2025
[3]

Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. 2022. M6-rec: Generative pretrained language models are open-ended recommender systems.arXiv preprint arXiv:2205.08084(2022)

work page arXiv 2022
[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 preprint arXiv:2502.18965 (2025)

work page internal anchor Pith review arXiv 2025
[5]

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. 299–315

2022
[6]

Mingzhe Han, Jiahao Liu, Dongsheng Li, Hansu Gu, Peng Zhang, Ning Gu, and Tun Lu. 2026. Feature-Indexed Federated Recommendation with Residual- Quantized Codebooks.arXiv preprint arXiv:2601.18570(2026)

work page arXiv 2026
[7]

Minjie Hong, Yan Xia, Zehan Wang, Jieming Zhu, Ye Wang, Sihang Cai, Xi- aoda Yang, Quanyu Dai, Zhenhua Dong, Zhimeng Zhang, et al. 2025. EAGER- LLM: Enhancing Large Language Models as Recommenders through Exogenous Behavior-Semantic Integration. InProceedings of the ACM on Web Conference

2025
[8]

Yupeng Hou, Jiacheng Li, Ashley Shin, Jinsung Jeon, Abhishek Santhanam, Wei Shao, Kaveh Hassani, Ning Yao, and Julian McAuley. 2025. Generating long semantic ids in parallel for recommendation. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 956–966

2025
[9]

Zhaoyu Hu, Jianyang Wang, Hao Guo, Yuan Tian, Erpeng Xue, Xianyang Qi, Hongxiang Lin, Lei Wang, and Sheng Chen. 2025. Dynamic Forgetting and Spatio- Temporal Periodic Interest Modeling for Local-Life Service Recommendation. arXiv preprint arXiv:2508.02451(2025)

work page arXiv 2025
[10]

Hao Jiang, Guoquan Wang, Sheng Yu, Yang Zeng, Wencong Zeng, and Guorui Zhou. 2025. A Plug-and-Play Spatially-Constrained Representation Enhancement Framework for Local-Life Recommendation.arXiv preprint arXiv:2511.12947 (2025)

work page internal anchor Pith review Pith/arXiv arXiv 2025
[11]

Hao Jiang, Guoquan Wang, Donglin Zhou, Sheng Yu, Yang Zeng, Wencong Zeng, Kun Gai, and Guorui Zhou. 2025. Llm-aligned geographic item tokenization for local-life recommendation.arXiv preprint arXiv:2511.14221(2025)

work page arXiv 2025
[12]

Xiaopeng Li, Bo Chen, Junda She, Shiteng Cao, You Wang, Qinlin Jia, Haiying He, Zheli Zhou, Zhao Liu, Ji Liu, et al. 2025. A survey of generative recommendation from a tri-decoupled perspective: Tokenization, architecture, and optimization. (2025)

2025
[13]

Defu Lian, Cong Zhao, Xing Xie, Guangzhong Sun, Enhong Chen, and Yong Rui
[14]

InProceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining

GeoMF: joint geographical modeling and matrix factorization for point-of- interest recommendation. InProceedings of the 20th ACM SIGKDD international conference on Knowledge discovery and data mining. 831–840
[15]

Nicholas Lim, Bryan Hooi, See-Kiong Ng, Xueou Wang, Yong Liang Goh, Ren- rong Weng, and Jagannadan Varadarajan. 2020. STP-UDGAT: Spatial-temporal- preference user dimensional graph attention network for next POI recommenda- tion. InProceedings of the 29th ACM international conference on information & knowledge management. 845–854

2020
[16]

Zhanyu Liu, Shiyao Wang, Xingmei Wang, Rongzhou Zhang, Jiaxin Deng, Honghui Bao, Jinghao Zhang, Wuchao Li, Pengfei Zheng, Xiangyu Wu, et al
[17]

Onerec-think: In-text reasoning for generative recommendation.arXiv preprint arXiv:2510.11639(2025)

work page arXiv 2025
[18]

Shashank Rajput, Nikhil Mehta, Anima Singh, Raghunandan Hulikal Keshavan, Trung Vu, Lukasz Heldt, Lichan Hong, Yi Tay, Vinh Tran, Jonah Samost, et al
[19]

Recommender systems with generative retrieval.Advances in Neural Information Processing Systems36 (2023), 10299–10315

2023
[20]

Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. 2024. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing568 (2024), 127063

2024
[21]

Aaron Van Den Oord, Oriol Vinyals, et al. 2017. Neural discrete representation learning.Advances in neural information processing systems30 (2017)

2017
[22]

Dongsheng Wang, Yuxi Huang, Shen Gao, Yifan Wang, Chengrui Huang, and Shuo Shang. 2025. Generative next poi recommendation with semantic id. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2. 2904–2914

2025
[23]

Guoquan Wang, Qiang Luo, Weisong Hu, Pengfei Yao, Wencong Zeng, Guorui Zhou, and Kun Gai. 2025. FIM: Frequency-aware multi-view interest modeling for local-life service recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1748–1757

2025
[24]

Shijia Wang, Tianpei Ouyang, Qiang Xiao, Dongjing Wang, Yintao Ren, Song- pei Xu, Da Guo, and Chuanjiang Luo. 2025. Progressive Semantic Residual Quantization for Multimodal-Joint Interest Modeling in Music Recommenda- tion. InProceedings of the 34th ACM International Conference on Information and Knowledge Management. 6119–6127

2025
[25]

Yejing Wang, Shengyu Zhou, Jinyu Lu, Qidong Liu, Xinhang Li, Wenlin Zhang, Feng Li, Pengjie Wang, Jian Xu, Bo Zheng, et al. 2025. GFlowGR: Fine-tuning Generative Recommendation Frameworks with Generative Flow Networks.arXiv preprint arXiv:2506.16114(2025)

work page arXiv 2025
[26]

Zhipeng Wei, Kuo Cai, Junda She, Jie Chen, Minghao Chen, Yang Zeng, Qiang Luo, Wencong Zeng, Ruiming Tang, Kun Gai, et al. 2025. Oneloc: Geo-aware generative recommender systems for local life service.arXiv preprint arXiv:2508.14646(2025)

work page arXiv 2025
[27]

An Yang, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoyan Huang, Jiandong Jiang, Jianhong Tu, Jianwei Zhang, Jingren Zhou, et al. 2025. Qwen2. 5-1m technical report.arXiv preprint arXiv:2501.15383(2025)

work page internal anchor Pith review arXiv 2025
[28]

Dingqi Yang, Daqing Zhang, Vincent W Zheng, and Zhiyong Yu. 2014. Modeling user activity preference by leveraging user spatial temporal characteristics in LBSNs.IEEE Transactions on Systems, Man, and Cybernetics: Systems45, 1 (2014), 129–142

2014
[29]

Yuhao Yang, Zhi Ji, Zhaopeng Li, Yi Li, Zhonglin Mo, Yue Ding, Kai Chen, Zijian Zhang, Jie Li, Shuanglong Li, et al. 2025. Sparse meets dense: Unified generative recommendations with cascaded sparse-dense representations.arXiv preprint arXiv:2503.02453(2025)

work page arXiv 2025
[30]

Jun Zhang, Yi Li, Yue Liu, Changping Wang, Yuan Wang, Yuling Xiong, Xun Liu, Haiyang Wu, Qian Li, Enming Zhang, et al. 2025. GPR: Towards a Generative Pre-trained One-Model Paradigm for Large-Scale Advertising Recommendation. arXiv preprint arXiv:2511.10138(2025). Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Trovato et al

work page arXiv 2025
[31]

Junjie Zhang, Beichen Zhang, Wenqi Sun, Hongyu Lu, Wayne Xin Zhao, Yu Chen, and Ji-Rong Wen. 2025. Slow Thinking for Sequential Recommendation.arXiv preprint arXiv:2504.09627(2025)

work page arXiv 2025
[32]

Yanzhao Zhang, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, Baosong Yang, Pengjun Xie, An Yang, Dayiheng Liu, Junyang Lin, et al. 2025. Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.arXiv preprint arXiv:2506.05176(2025)

work page internal anchor Pith review arXiv 2025
[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. In2024 IEEE 40th International Conference on Data Engineering (ICDE). IEEE, 1435–1448

2024
[34]

Guorui Zhou, Hengrui Hu, Hongtao Cheng, Huanjie Wang, Jiaxin Deng, Jinghao Zhang, Kuo Cai, Lejian Ren, Lu Ren, Liao Yu, et al. 2025. Onerec-v2 technical report.arXiv preprint arXiv:2508.20900(2025)

work page arXiv 2025
[35]

Jingyi Zhou, Cheng Chen, Kai Zuo, Manjie Xu, Zhendong Fu, Yibo Chen, Xu Tang, and Yao Hu. 2025. HyMiRec: A Hybrid Multi-interest Learning Framework for LLM-based Sequential Recommendation.arXiv preprint arXiv:2510.13738 (2025). A Proof of Equation(10) A.1 Notations and Preliminaries For ease of presentation, we introduce the following definitions. Definit...

work page arXiv 2025