Pith. sign in

REVIEW 2 major objections 6 minor 61 references

A hypergraph recommender that repairs missing POI content layer by layer beats 15 baselines on three location datasets.

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

2026-08-04 19:41 UTC pith:TI4GIBBH

load-bearing objection Solid, novel mechanism for coupling completion with hypergraph incidence refinement, but the layer-wise claim is conditional on an underspecified budget match and missing code. the 2 major comments →

arxiv 2608.01846 v1 pith:TI4GIBBH submitted 2026-08-03 cs.IR

HyperAgent4POI: Dynamic Semantic Message Passing on Multi-Agent Hypergraphs for Missing-Modality Recommendation

Jinze Wang , Yuze Liu , Tiehua Zhang , Jiong Jin , Zhu Sun This is my paper
classification cs.IR
keywords point-of-interest recommendationmissing modalitieshypergraph neural networksdynamic semantic message passingmodality completionsoft incidence refinementLLM-augmented recommendationnext POI ranking
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.

HyperAgent4POI tackles next point-of-interest recommendation when venue text or images are missing. Its central claim is that modality completion and hyperedge-topology refinement should happen together inside every message-passing layer, not once before propagation. On three LBSN benchmarks and across modality-missing rates from 20% to 80%, it reports consistent NDCG@20 gains, averaging 8.2% over the strongest baseline at 60% missingness. The method also keeps inference fast by caching final node representations so no LLM calls are needed at serving time.

Core claim

The paper proposes Dynamic Semantic Message Passing (DSMP), a layer-wise loop in which a persistent agent for each node produces semantic messages, these messages form a 'semantic hyperedge motif' per edge, and that motif is used to refine soft incidence weights and to complete missing modality features before the next layer. By redoing completion after each incidence update, DSMP prevents representation discrepancies from propagating unchanged through the hypergraph. The authors show that this layer-wise coupling improves both ranking and representation quality relative to one-shot pre-completion, with smaller performance degradation as missingness increases.

What carries the argument

The central mechanism is the DSMP loop composed of four role-conditioned operators—Local Expert, Semantic Aggregator, Topology Evolution, and Modality Deduction—that share a frozen large language model backbone with lightweight role-specific adapters. Local Expert converts node states into edge-specific messages; Semantic Aggregator summarizes them into a semantic hyperedge motif; Topology Evolution scores candidate POIs against that motif to produce bounded soft incidences; Modality Deduction reconstructs missing features using the refined hyperedge context. The completed features then update node states for the next layer, so completion and topology refinement are jointly conditioned on th

Load-bearing premise

The measured gains are attributed specifically to the layer-wise coupling of completion and incidence refinement, but the comparisons never isolate that coupling from the much larger frozen-LLM backbone and other joint design choices.

What would settle it

A controlled experiment that keeps the identical frozen backbone, adapters, training budget, and candidate-pool logic but replaces the per-layer DSMP loop with a single pre-propagation completion—matching the same total compute—would refute the central claim if it did not show a significant NDCG@20 drop across all modality-missing rates.

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the claim holds, next-POI recommenders can maintain ranking quality when venue text or photos are only partially available, instead of degrading sharply as modality coverage drops.
  • The layer-wise coupling provides a template for other graph-learning tasks with missing node features: completion should be conditioned on the evolving message-passing context rather than fixed upfront.
  • Cached node representations mean the expensive LLM-based semantic computation stays offline, making the approach practical for online serving despite a large frozen backbone.
  • The reported robustness trend—smaller relative NDCG decline at higher missingness—suggests the method becomes more valuable exactly when less modality evidence is present.
  • The differentiable soft incidence refinement with a conservative prior offers a path toward interpretable hypergraph topology: the exported hard incidences are thresholds of learned soft scores.

Where Pith is reading between the lines

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

  • Editorial: The DSMP principle likely transfers beyond POI recommendation to any recommendation task with heterogeneous item content that is partially observed, such as news, video, or e-commerce, where missing modalities are common.
  • Editorial: The paper's diagnostic of 'representation discrepancy' could be developed into a general measurable criterion—tracking how far a masked representation diverges from its complete counterpart during propagation—and used to decide where in a graph model to insert completion modules.
  • Editorial: A testable extension is to ask whether the layer-wise benefit increases with hypergraph diameter or with the number of co-visit POIs per edge; one might expect deeper propagation to amplify discrepancy, making DSMP even more useful on denser hypergraphs.
  • Editorial: Because the method relies on a frozen LLM backbone, its practical ceiling depends on the quality of that backbone's hidden states; with a stronger backbone the same mechanism could yield larger gains, while with a weaker one the gains might shrink.

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

2 major / 6 minor

Summary. HyperAgent4POI addresses next-POI recommendation when text/image modalities are missing. It builds an initial user–POI session hypergraph and, over L=2 DSMP layers, alternates semantic aggregation with a frozen Llama-3-8B backbone plus role-specific LoRA adapters, bounded soft-incidence refinement via a frozen retriever and compatible/incompatible scoring, and modality completion with a contrastive loss. Node representations are cached for dot-product serving. Experiments on Yelp-2018, FSQ-NYC, and FSQ-TKY compare against 15 baselines under random modality missingness at ρ_miss=0.6, reporting NDCG@20 gains of 7.4%–9.4% over HIRE and consistent gains across ρ_miss ∈ {0.2, 0.4, 0.6, 0.8}. RQ4 compares NoComp, PreComp, and Layerwise DSMP and reports a 4.5% mean NDCG@20 improvement of layerwise over pre-completion under a 'matched computational budget.' The paper's central claim is that layer-wise coupling of completion and incidence refinement, rather than scale or architecture alone, drives the gains.

Significance. If substantiated, the contribution is valuable: it is a concrete instantiation of completing missing modalities inside hypergraph message passing, with an effective caching scheme for LLM-free serving. The masking protocol cleanly separates completion targets from model inputs, and the experimental setup spans three datasets, 15 baselines, multiple missing rates, and a reasonable set of ablations. Credit is due for the RQ4 diagnostic (NoComp/PreComp/Layerwise) and for A6, which includes a parameter-matched MLP ablation rather than only dropping components. The paper is not circular: the ranking objective is BPR on held-out next POIs, while the completion loss is supervised on masked targets. The open issue is causal isolation of the layer-wise DSMP loop from compute, backbone, and system-level confounds; the current evidence is suggestive but not yet compelling.

major comments (2)
  1. [RQ4: Layer-wise Completion (Table 4)] The central claim that layer-wise completion is beneficial rests on the 4.5% NDCG@20 improvement of Layerwise DSMP over PreComp, but the phrase 'matched computational budget' is not defined. Given the main model has L=2, Layerwise performs two Modality Deduction calls per POI while PreComp performs one. If the budget is wall-clock time or parameter count only, the gain may reflect additional LLM forward passes rather than the layer-wise coupling the paper asserts. Please specify exactly how the budget is equalized (e.g., equal LLM calls, FLOPs, or total training time), and report per-dataset NDCG@20 values with standard deviations; Table 4 currently reports only the mean across three datasets, and the 4.5% could be driven by one dataset.
  2. [RQ1: Main Performance Comparison (Table 2)] The headline gains over HIRE (7.4%, 9.4%, and 7.8% NDCG@20 at ρ_miss=0.6) are cross-system comparisons: HIRE does not use the frozen Llama-3-8B backbone, the LoRA role-conditioned operators, or the same auxiliary losses. A6 (Figure 3) matches parameter count against residual MLPs but does not control for pretrained knowledge or the other system-level differences. Thus Table 2 does not isolate the marginal effect of DSMP's layer-wise loop. A controlled variant of HIRE (or the strongest baseline) with the same backbone, completion module, and losses but a one-shot schedule, or an explicit analysis built on fully specified RQ4 statistics, is needed before attributing the margin to layer-wise coupling.
minor comments (6)
  1. [Figure 3] The ablation figure labels contain garbled '/uni000000...' token sequences in the provided version. Please replace them with the actual dataset and method labels.
  2. [RQ4 diagnostics] CompCos is mentioned as a diagnostic but never defined; please provide its exact formula. Also, CompCos and RepDisc are computed only on Yelp-2018 with no error bars; please state whether the reported values are single-seed or averaged and add variance information.
  3. [Methodology, Equation (1)] Equation (1) is a layer-wise recurrence for a static linear hypergraph operator. It is fine as motivation for 'discrepancy propagation,' but the text should not be read as evidence that layer-wise completion is necessary; the actual evidence must come from RQ4. Consider one clarifying sentence to that effect.
  4. [Contributions] Contribution 1 says the paper 'formalizes incomplete multimodal next POI recommendation,' but the content is a problem formulation plus a motivating linearization. 'Formulates' or 'frames' would be more accurate than 'formalizes.'
  5. [General] No code availability statement is provided. Since the method involves custom prompt templates, frozen encoders, and several loss coefficients, a reproducibility statement or a note on whether code will be released would help.
  6. [RQ2, Table 3] The robustness table reports only mean NDCG@20 over five seeds, with no standard deviations or significance tests for the trend. Adding error bars would make the 'smaller relative decline' claim easier to verify.

Circularity Check

0 steps flagged

No significant circularity: ranking and completion objectives are grounded in external held-out targets; DSMP is an iterative architectural claim, not a definitional tautology.

full rationale

The paper's central claim is that coupling modality completion with soft incidence refinement inside each hypergraph layer improves next-POI ranking. I walked the derivation chain: node states are initialized from observed features and a learned missing-modality embedding (Eq. 2); messages and motifs are computed from these states (Eq. 3); Topology Evolution produces soft incidences from teacher-forced LLM label likelihoods (Eq. 4); Modality Deduction generates completed features from refined hyperedge context (Eq. 5). The ranking objective (Eq. 6) is BPR on held-out next-POI triples, and the completion loss (Eq. 7) is a contrastive objective against true masked encoder features z_j^k, which are excluded from prompts and inputs. Thus neither the ranking target nor the completion target is defined in terms of the model's own output; the loop is an iterative refinement, not a tautology. The linearized discrepancy-propagation equation (Eq. 1) is a motivating algebraic identity, not a predictive claim, so it does not constitute a circular step. The paper does cite the authors' prior work for dataset protocols (Wang et al. 2025b, 2026a) and related LLM-agent designs, but those citations are not load-bearing: the experimental evaluation against 15 external baselines and the held-out ranking protocol provide independent content. The remaining weaknesses—e.g., the underspecified 'matched computational budget' in RQ4 and the training-set CompCos diagnostic—are soundness/control concerns, not circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 2 invented entities

The paper introduces one new model with a rich set of trainable components. The main load-bearing assumptions are the efficacy of the joint DSMP loop and the quality of the external data matching. Most free parameters are standard hyperparameters, but their values matter for replication.

free parameters (8)
  • modality-missing rate rho_miss = {0.2,0.4,0.6,0.8}
    This is a simulation parameter that controls the experimental setup, not a fitted constant of the model.
  • loss coefficients lambda and gamma = lambda=0.5, gamma=0.1
    Joint weights for the auxiliary completion and topology terms, chosen by validation.
  • export threshold tau = 0.75
    Threshold for binarizing soft incidences.
  • retrieval budget K_ret = 5
    Number of top candidates retrieved for the topology evolution pool.
  • LoRA rank r and scaling alpha = 16 and 32
    LoRA adapter hyperparameters.
  • temperature T = 0.07
    Temperature in the InfoNCE completion loss.
  • epsilon_numerical and epsilon_lab = 1e-6
    Numerical stabilizers in normalization and in the topology KL loss.
  • number of DSMP layers L = 2
    The model uses two DSMP layers. The effect of L is not studied in the main text.
axioms (4)
  • domain assumption Soft incidence refinement and modality completion inside a shared propagation loop can be jointly optimized without degrading each other.
    This is the central design premise. The paper gives an example in the motivating linearization, but does not prove convergence or stability of the joint loop.
  • domain assumption Foursquare check-ins can be matched to Google Places images and Yelp reviews with 96.8% accepted-match precision.
    The FSQ-NYC and FSQ-TKY experiments depend on this matching, which is described only in the supplementary material.
  • domain assumption The frozen Llama-3-8B backbone is a sufficient semantic operator for role-conditioned messages.
    The model's performance depends on the frozen-LLM hidden states carrying the semantics needed for the four roles, which is validated mainly through ablation.
  • domain assumption POI text and image features encoded with frozen MiniLM and CLIP are informative enough for modality completion.
    The completed features are trained to match these encodings. If the encodings are weak, the completion targets are weak.
invented entities (2)
  • persistent node agents with role-specific LoRA adapters no independent evidence
    purpose: To produce node-to-hyperedge messages, semantic hyperedge motifs, incidence scores, and completed modality features.
    These are internal architectural constructs without external falsifiable handles. The entities are not claimed to exist outside the model.
  • semantic hyperedge motifs no independent evidence
    purpose: To summarize node evidence for incidence refinement and modality completion.
    Defined only as aggregated hidden states. Their semantic content is asserted but not independently verified.

pith-pipeline@v1.3.0-daily-deepseek · 3494 in / 4524 out tokens · 83381 ms · 2026-08-04T19:41:10.401136+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of HyperAgent4POI: Dynamic Semantic Message Passing on Multi-Agent Hypergraphs for Missing-Modality Recommendation." pith.science (2026). https://pith.science/paper/TI4GIBBH

@misc{pith2026260801846,
  author       = {Pith},
  title        = {Pith review of: HyperAgent4POI: Dynamic Semantic Message Passing on Multi-Agent Hypergraphs for Missing-Modality Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TI4GIBBH}},
  note         = {Machine review of arXiv:2608.01846}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Next Point-of-Interest (POI) recommendation benefits from textual and visual content that describes venue semantics, yet such content is often incomplete in real-world services. Missing modalities weaken POI representations and reduce the semantic evidence available for ranking. The resulting representations also provide unreliable evidence for modeling higher-order user--POI interactions. We propose HyperAgent4POI, which uses Dynamic Semantic Message Passing (DSMP) to perform modality completion and soft incidence refinement within each hypergraph layer. Persistent node agents share a frozen Llama backbone and use role-specific adapters to produce node-to-hyperedge messages. Semantic hyperedge motifs formed from these messages guide soft incidence scoring and modality completion. Final node representations are cached for online ranking without LLM calls. Experiments on three real-world LBSN datasets show consistent ranking gains over 15 baselines across modality-missing rates, while cached inference provides practical online efficiency. Under a 60% modality-missing rate, HyperAgent4POI improves NDCG@20 over the strongest baseline by 8.2% on average across the three datasets.

Figures

Figures reproduced from arXiv: 2608.01846 by Jinze Wang, Jiong Jin, Tiehua Zhang, Yuze Liu, Zhu Sun.

Figure 1
Figure 1. Figure 1: Overall architecture of HyperAgent4POI and its layer-wise DSMP workflow. Each DSMP layer uses role-conditioned [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Prompt templates for the four role-conditioned se [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Ablation on Yelp-2018 (ρmiss = 0.6) over five seeds; labels denote NDCG@20 reductions from the full model. increases from 0.2 to 0.8, NDCG@20 decreases by 25.6– 26.7% for HIRE and by 17.3–18.2% for HyperAgent4POI. At ρmiss = 0.8, the performance gap widens, while HyperA￾gent4POI shows a smaller relative decline. This result sug￾gests that layer-wise contextual completion becomes increas￾ingly beneficial wh… 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

61 extracted references · 7 linked inside Pith

  1. [1]

    AAAI , volume=

    A survey of point-of-interest recommendation in location-based social networks , author=. AAAI , volume=

  2. [2]

    IEEE TKDE , volume=

    Where to go next: A spatio-temporal gated network for next poi recommendation , author=. IEEE TKDE , volume=. 2020 , publisher=

    2020

  3. [3]

    ACM TOIS , volume=

    The footprint of factorization models and their applications in collaborative filtering , author=. ACM TOIS , volume=. 2021 , publisher=

    2021

  4. [4]

    ACM TOIS , volume=

    Unified relevance models for rating prediction in collaborative filtering , author=. ACM TOIS , volume=. 2008 , publisher=

    2008

  5. [5]

    SIGIR , pages=

    Exploiting geographical influence for collaborative point-of-interest recommendation , author=. SIGIR , pages=

  6. [6]

    AAAI , volume=

    Where to go next: Modeling long-and short-term user preferences for point-of-interest recommendation , author=. AAAI , volume=

  7. [7]

    CIKM , pages=

    Graph-based point-of-interest recommendation with geographical and temporal influences , author=. CIKM , pages=

  8. [8]

    ICDM , pages=

    STGCN: a spatial-temporal aware graph learning method for POI recommendation , author=. ICDM , pages=. 2020 , organization=

    2020

  9. [9]

    SIGIR , pages=

    Spatio-temporal hypergraph learning for next POI recommendation , author=. SIGIR , pages=

  10. [10]

    KDD , pages=

    Are we really making much progress? revisiting, benchmarking and refining heterogeneous graph neural networks , author=. KDD , pages=

  11. [11]

    SIGIR , pages=

    Hypergraph contrastive collaborative filtering , author=. SIGIR , pages=

  12. [12]

    KDD , pages=

    Dual channel hypergraph collaborative filtering , author=. KDD , pages=

  13. [13]

    WSDM , pages=

    When Modalities Go Missing: Early Fusion for Multimodal Recommendation , author=. WSDM , pages=

  14. [14]

    SIGIR , pages=

    Disentangling and generating modalities for recommendation in missing modality scenarios , author=. SIGIR , pages=

  15. [15]

    WWW , pages=

    Generating with fairness: A modality-diffused counterfactual framework for incomplete multimodal recommendations , author=. WWW , pages=

  16. [16]

    KBS , volume=

    Graph attention contrastive learning with missing modality for multimodal recommendation , author=. KBS , volume=. 2025 , publisher=

    2025

  17. [17]

    ACM TOIS , volume=

    Unified Heterogeneous Hypergraph Construction for Incomplete Multimedia Recommendation , author=. ACM TOIS , volume=. 2025 , publisher=

    2025

  18. [18]

    KDD , pages=

    Hgmf: heterogeneous graph-based fusion for multimodal data with incompleteness , author=. KDD , pages=

  19. [19]

    ICLR , volume=

    Learnable expansion of graph operators for multi-modal feature fusion , author=. ICLR , volume=

  20. [20]

    ICCV , pages=

    Missrag: Addressing the missing modality challenge in multimodal large language models , author=. ICCV , pages=

  21. [21]

    AAAI , volume=

    Generating-Filtering-Ranking: A Three-Stage MultiModal Data Augmentation Framework Under Partial Modality Missing , author=. AAAI , volume=

  22. [22]

    ACM TIST , volume=

    Participatory cultural mapping based on collective behavior data in location-based social networks , author=. ACM TIST , volume=. 2016 , publisher=

    2016

  23. [23]

    CIKM , pages=

    The Yelp Collaborative Knowledge Graph , author=. CIKM , pages=

  24. [24]

    arXiv preprint arXiv:1205.2618 , year=

    BPR: Bayesian personalized ranking from implicit feedback , author=. arXiv preprint arXiv:1205.2618 , year=

    Pith/arXiv arXiv

  25. [25]

    SIGIR , pages=

    Neural graph collaborative filtering , author=. SIGIR , pages=

  26. [26]

    SIGIR , pages=

    Self-supervised graph learning for recommendation , author=. SIGIR , pages=

  27. [27]

    SIGIR , pages=

    Are graph augmentations necessary? simple graph contrastive learning for recommendation , author=. SIGIR , pages=

  28. [28]

    CIKM , pages=

    Dynamic hypergraph learning for collaborative filtering , author=. CIKM , pages=

  29. [29]

    CIKM , pages=

    Double-scale self-supervised hypergraph learning for group recommendation , author=. CIKM , pages=

  30. [30]

    IJCAI , pages=

    Beyond Individual and Point: Next POI Recommendation via Region-aware Dynamic Hypergraph with Dual-level Modeling , author=. IJCAI , pages=

  31. [31]

    Findings of NAACL , pages=

    Llm-rec: Personalized recommendation via prompting large language models , author=. Findings of NAACL , pages=

  32. [32]

    SIGIR , pages=

    MSCRS: Multi-modal semantic graph prompt learning framework for conversational recommender systems , author=. SIGIR , pages=

  33. [33]

    WWW , pages=

    Representation learning with large language models for recommendation , author=. WWW , pages=

  34. [34]

    WWW , pages=

    Anchor drift no more: Hierarchical consistency-guided prompt distillation for incomplete multimodal learning , author=. WWW , pages=

  35. [35]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. ICLR , volume=

  36. [36]

    NeurIPS , volume=

    Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers , author=. NeurIPS , volume=

  37. [37]

    EMNLP , pages=

    LRMM: Learning to recommend with missing modalities , author=. EMNLP , pages=

  38. [38]

    RecSys , pages=

    A multimodal single-branch embedding network for recommendation in cold-start and missing modality scenarios , author=. RecSys , pages=

  39. [39]

    arXiv preprint arXiv:2605.00670 , year=

    Robust Multimodal Recommendation via Graph Retrieval-Enhanced Modality Completion , author=. arXiv preprint arXiv:2605.00670 , year=

    Pith/arXiv arXiv

  40. [40]

    EMNLP , year=

    Lagcl4rec: When llms activate interactions potential in graph contrastive learning for recommendation , author=. EMNLP , year=

  41. [41]

    EMNLP , pages=

    Enhancing high-order interaction awareness in llm-based recommender model , author=. EMNLP , pages=

  42. [42]

    arXiv preprint arXiv:2401.08217 , year=

    Llm-guided multi-view hypergraph learning for human-centric explainable recommendation , author=. arXiv preprint arXiv:2401.08217 , year=

    Pith/arXiv arXiv

  43. [43]

    arXiv preprint arXiv:2504.10541 , year=

    Multi-Modal Hypergraph Enhanced LLM Learning for Recommendation , author=. arXiv preprint arXiv:2504.10541 , year=

    Pith/arXiv arXiv

  44. [44]

    SIGIR , pages=

    Macrec: A multi-agent collaboration framework for recommendation , author=. SIGIR , pages=

  45. [45]

    arXiv preprint arXiv:2502.10050 , year=

    A survey on llm-powered agents for recommender systems , author=. arXiv preprint arXiv:2502.10050 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2510.08012 , year=

    Do we really need sft? prompt-as-policy over knowledge graphs for cold-start next poi recommendation , author=. arXiv preprint arXiv:2510.08012 , year=

    arXiv

  47. [47]

    NAACL , pages=

    A Multi-modal Large Language Model with Graph-of-Thought for Effective Recommendation , author=. NAACL , pages=

  48. [48]

    RecSys , pages=

    Vl-clip: Enhancing multimodal recommendations via visual grounding and llm-augmented clip embeddings , author=. RecSys , pages=

  49. [49]

    FGCS , volume=

    Graph convolutional networks for graphs containing missing features , author=. FGCS , volume=. 2021 , publisher=

    2021

  50. [50]

    LoG , pages=

    On the unreasonable effectiveness of feature propagation in learning on graphs with missing node features , author=. LoG , pages=. 2022 , organization=

    2022

  51. [51]

    DMKD , volume=

    Large language models are zero-shot point-of-interest recommenders , author=. DMKD , volume=. 2025 , publisher=

    2025

  52. [52]

    arXiv preprint arXiv:2506.21599 , year=

    Refine-POI: Reinforcement Fine-Tuned Large Language Models for Next Point-of-Interest Recommendation , author=. arXiv preprint arXiv:2506.21599 , year=

    arXiv

  53. [53]

    arXiv preprint arXiv:2605.15203 , year=

    Agent4POI: Agentic Context-Conditioned Affordance Reasoning for Multimodal Point-of-Interest Recommendation , author=. arXiv preprint arXiv:2605.15203 , year=

    Pith/arXiv arXiv

  54. [54]

    DASFAA , pages=

    Multi-view spatial-temporal enhanced hypergraph network for next poi recommendation , author=. DASFAA , pages=. 2023 , organization=

    2023

  55. [55]

    arXiv preprint arXiv:2601.11610 , year=

    Multifaceted Scenario-Aware Hypergraph Learning for Next POI Recommendation , author=. arXiv preprint arXiv:2601.11610 , year=

    arXiv

  56. [56]

    IEEE TSC , year=

    Hgdrec: Next poi recommendation based on hypergraph neural network and diffusion model , author=. IEEE TSC , year=

  57. [57]

    IPM , volume=

    Global and local hypergraph learning method with semantic enhancement for POI recommendation , author=. IPM , volume=. 2025 , publisher=

    2025

  58. [58]

    arXiv preprint arXiv:2605.25007 , year=

    Meta-Modal Agent: Sequential Evidence Routing for Missing-Modality Candidate Reranking , author=. arXiv preprint arXiv:2605.25007 , year=

    Pith/arXiv arXiv

  59. [59]

    ICME , pages=

    HyperMAN: Hypergraph-enhanced Meta-learning Adaptive Network for Next POI Recommendation , author=. ICME , pages=. 2025 , organization=

    2025

  60. [60]

    PAKDD , pages=

    Meta-learning enhanced next POI recommendation by leveraging check-ins from auxiliary cities , author=. PAKDD , pages=. 2023 , organization=

    2023

  61. [61]

    ICML , pages=

    Learning transferable visual models from natural language supervision , author=. ICML , pages=. 2021 , organization=

    2021