Source-linked AI summary
Category-Anchored Large Language Models for Next POI Recommendation
Lexi Bai, Xiaopeng Zeng, Xun Zhou, Yuanshao Zhu, James Jianqiao Yu
TL;DR
Next POI recommendation asks models to predict a concrete destination from check-in histories, but direct LLM generation faces a semantic-to-entity gap in sparse, localized POI spaces. ANCHOR estimates category-level intent, grounds candidate POIs using category, history, and preference evidence, and reranks them with historical and contextual evidence. It outperforms the strongest LLM-based baseline Think2Go on Acc@1 and MRR across three datasets, while a stated limitation is reliance on sufficient mobility history.
Problem
LLMs can infer mobility intent, but predicting concrete POIs from sparse, localized, long-tailed entity spaces creates a semantic-to-entity gap.
Source: §I · p. 1Method
ANCHOR predicts category-level intent, grounds POIs using category, history, and preference evidence, then reranks candidates using historical mobility patterns and context.
Source: §I · p. 2Results
ANCHOR improves Acc@1 over Think2Go by 7.56%, 11.21%, and 1.94% on NYC, TKY, and CA, and MRR by 14.41%, 12.66%, and 8.37%, respectively.
Source: B. Main Results · p. 8Takeaways & Limitations
Across three real-world datasets, ANCHOR reports favorable trade-offs among candidate coverage, ranking quality, and inference efficiency.
Source: §I · p. 2ANCHOR depends on sufficient historical mobility data for grounding and reranking, which may limit robustness for extreme cold-start users with very short histories.
Source: §VI · p. 12
Abstract
from arXiv · showhide
Next point-of-interest (POI) recommendation aims to predict a user's next destination from spatio-temporal check-in data, and serves as the foundation for many location-based services. Recent advances incorporate large language models (LLMs) into this task by textualizing user trajectories to exploit their semantic reasoning ability. However, directly generating POI identifiers with such models is unreliable: POIs are sparse, localized, and long-tailed entities with weak natural-language semantics, which creates a semantic-to-entity gap between high-level mobility intent and concrete POI prediction. To bridge this gap, we propose ANCHOR, a category-anchored framework for semantic-to-entity grounding in next POI recommendation. Specifically, ANCHOR first estimates category-level mobility intent in a compact and stable semantic space, rather than predicting concrete POI entities directly. The inferred intent is then grounded into a candidate set by integrating three complementary sources of mobility evidence: category anchors, historical visit memory, and user preferences. Finally, a history-aware reranker selects the target POI by matching each candidate with the user's historical mobility patterns and current context. Experiments on three real-world check-in datasets show that ANCHOR consistently outperforms representative traditional, structure-enhanced, and LLM-based baselines on the reported metrics. Further analyses validate the effectiveness of semantic-to-entity grounding and show favorable trade-offs among candidate coverage, ranking quality, and inference efficiency.
I. INTRODUCTION
Next POI recommendation must infer a concrete destination from sparse, heterogeneous check-in histories, while LLMs’ high-level mobility reasoning does not directly ensure reliable entity prediction. ANCHOR bridges this semantic-to-entity gap by predicting category-level intent, grounding candidates with multiple evidence sources, and reranking them using historical and contextual evidence.
I. INTRODUCTION
- Directly generating POI identifiers is unreliable because POIs occupy sparse, localized, dataset-specific spaces with weak and unstable semantics.The resulting mismatch between high-level mobility intent and concrete entity prediction is the semantic-to-entity gap.Source: §I · pp. 1–2
- Reliable prediction must address symbolic POI identifiers, one-to-many ambiguity from coarse intent, and conflicts among context-dependent evidence sources.Relevant evidence can include user history, temporal context, spatial proximity, and structural relations.Source: §I · p. 2
- ANCHOR predicts category-level mobility intent, grounds it using category, history, and preference evidence, then reranks candidates against historical mobility patterns and context.This reason–ground–discriminate sequence connects semantic reasoning to concrete POI prediction.Source: §I · p. 2
- ANCHOR consistently outperforms baseline models on three real-world datasets, with favorable trade-offs among candidate coverage, ranking quality, and inference efficiency.The supplied introduction reports the overall pattern but does not give metric values.Source: §I · p. 2
II. PRELIMINARIES
- Next POI recommendation ranks candidate POIs for a user’s historical trajectory and target timestamp.The task definition represents check-ins as time-ordered records containing POI, category, timestamp, and geographical coordinates.Source: §II · p. 2
III. METHODOLOGY
ANCHOR decomposes next-POI prediction into category-level intent estimation, multi-source candidate grounding, and history-aware reranking. Its LLM predicts stable category labels from compact trajectory prompts rather than generating POI identifiers.
A. Framework Overview
- ANCHOR separates prediction into semantic intent pre-estimation, multi-source candidate grounding, and history-aware reranking.The modules respectively predict category-level intent, construct the candidate POI pool, and produce the final recommendation.Source: A. Framework Overview · p. 2, Figure 2 · p. 3
B. Semantic Intent Pre-estimation via LLM
- The LLM estimates mobility intent in a compact category space instead of generating a concrete POI ID.Category classification restricts outputs to predefined POI categories, addressing the mismatch and uncontrollable outputs associated with direct POI generation.Source: A. Framework Overview · p. 2, B. Semantic Intent Pre-estimation via… · pp. 3–4
- Compact field-based prompts represent check-ins with spatio-temporal attributes to limit the token overhead of sentence-level textualization.The prompt uses instruction, history, current-context, and output blocks; its history records trajectory ID, time, day of week, normalized within-day time, and POI category.Source: B. Semantic Intent Pre-estimation via… · p. 3
- ANCHOR adapts the LLM with LoRA while keeping backbone parameters frozen and training adapters and the classification head.The category classifier is trained using cross-entropy on trajectory prompts and target categories.Source: B. Semantic Intent Pre-estimation via… · p. 4
C. Multi-Source Candidate Grounding
Category intent alone leaves ambiguity because many concrete POIs can share a category. ANCHOR grounds candidates using category evidence, visit history, and user preferences, then selects a balanced pool for reranking.
C. Multi-Source Candidate Grounding
- A category can contain many POIs with different historical relevance, preferences, and contextual fit, leaving ambiguity between category intent and the target entity.Relying only on predicted categories can miss the ground-truth POI when categories are incomplete or flood the pool with weakly related POIs when categories are broad.Source: C. Multi-Source Candidate Grounding · p. 4
- ANCHOR combines predicted-category POIs, historical POI memory, and user-preferred category expansion to form candidate sources.These sources respectively reflect semantic intent, revisits, and long-term, recent, and target-time preferences.Source: C. Multi-Source Candidate Grounding · p. 4
- Candidate grounding assigns a prior from LLM category evidence, visit history, and user preference; this prior organizes candidates but is not the final recommendation score.History evidence incorporates visit frequency and recency, while preference evidence reflects long-term, recent, and target-time category preferences.Source: C. Multi-Source Candidate Grounding · p. 5
- Source-specific quotas prevent the larger LLM-category set from crowding out smaller sources such as historical POI memory.ANCHOR ranks candidates within each source, merges and deduplicates quota selections, then fills remaining slots by grounding prior up to pool size N.Source: C. Multi-Source Candidate Grounding · p. 5
D. History-Aware Reranking
ANCHOR reranks grounded candidates by matching them against user history and current context. Candidate-specific attention emphasizes relevant past visits, while context-aware scoring combines historical, temporal, and transition signals.
D. History-Aware Reranking
- The reranker addresses the gap between source-level candidate support and fine-grained relevance to the user’s current state.It evaluates candidate fit with historical behavior, target-time conditions, and recent mobility state.Source: D. History-Aware Reranking · p. 5
- Each candidate attends to relevant historical check-ins, producing a candidate-specific historical context for reranking.History and candidate representations share a hidden space; the history representation fuses POI, category, time, and trajectory-position information.Source: D. History-Aware Reranking · pp. 5–6
- Three context queries represent recent historical state, target-time condition, and short-term transition state.ANCHOR uses the recent-history query to pool the context queries before matching the resulting prediction state with candidates.Source: D. History-Aware Reranking · p. 6
- Candidate scoring combines its history-aware representation with overall, target-time, and short-term transition context interactions.An MLP maps the combined representation to a ranking score; inference returns the highest-scored POI in the candidate pool.Source: D. History-Aware Reranking · p. 7
A. Experimental Settings
The experiments use three public check-in datasets, chronological trajectory-level splits, and Acc@K and MRR metrics to compare ANCHOR with representative recommendation baselines.
A. Experimental Settings
- NYC, TKY, and CA check-ins are filtered to retain POIs and users with at least 10 check-ins, then split chronologically 80/10/10 by trajectory.The datasets cover Foursquare activity in New York and Tokyo during 2012–2013 and Gowalla activity in California and Nevada during 2009–2010.Source: A. Experimental Settings · p. 7, Table III · p. 7
- Acc@1, Acc@5, and Acc@10 measure whether the true POI appears within the top K ranks, while MRR evaluates ranking quality.The evaluation uses ranked POI lists for each test sample.Source: A. Experimental Settings · p. 7
- The comparison evaluates ANCHOR against representative next-POI recommendation baselines spanning sequential, spatio-temporal, structure-aware, temporal, and LLM-based methods, including GETNext, STHGCN, ROTAN, LLM4POI, and GNPR.Source: A. Experimental Settings · p. 7
B. Main Results
ANCHOR performs best across NYC, TKY, and CA on Acc@1, Acc@5, Acc@10, and MRR. Its reported gains over non-LLM and LLM-based baselines are consistent across the three datasets.
B. Main Results
- ANCHOR achieves the best results on NYC, TKY, and CA across Acc@1, Acc@5, Acc@10, and MRR.The reported advantage spans different cities, platforms, and recommendation-list lengths.Source: B. Main Results · p. 8
- 33.23%, 46.54%, and 19.74% are ANCHOR’s Acc@1 relative gains over ROTAN on NYC, TKY, and CA, respectively.The paper notes that POI-level transition patterns alone may be insufficient when the entity space is sparse, long-tailed, and lacks stable semantics.Source: B. Main Results · p. 8
- 7.56%, 11.21%, and 1.94% are ANCHOR’s Acc@1 relative gains over Think2Go on NYC, TKY, and CA, respectively.ANCHOR also exceeds Think2Go on MRR by 14.41%, 12.66%, and 8.37% on NYC, TKY, and CA, respectively.Source: B. Main Results · p. 8
C. Ablation Study
The full ANCHOR framework outperforms its ablation variants on all three datasets, with the results attributing contributions to structural initialization, semantic intent, multi-source candidate generation, and reranking.
C. Ablation Study
- The complete ANCHOR framework consistently outperforms all ablation variants on NYC, TKY, and CA in Acc@1.The ablations replace or remove LightGCN initialization, the LLM anchor, multi-source grounding, or the history-aware reranker.Source: C. Ablation Study · pp. 8–9, Table V · p. 9
- Removing LightGCN initialization reduces performance, highlighting the benefit of structural transition patterns in POI representations before contextual reranking.Removing the LLM anchor also degrades performance; the authors attribute its role to constraining predictions beyond historical repetition.Source: C. Ablation Study · p. 9
- Restricting candidates to the LLM category anchor underperforms multi-source grounding, which adds history and preference signals to improve candidate-pool coverage and robustness.The authors also report that replacing the history-aware reranker with a standard MLP degrades performance because candidate recall alone is inadequate.Source: C. Ablation Study · p. 9
D. Effectiveness of Semantic Intent Pre-estimation
The analyses test category coverage as K grows and compare the fine-tuned LLM with alternative category predictors. The results link category-level prediction quality to final POI recommendation performance.
D. Effectiveness of Semantic Intent Pre-estimation
- Category Acc@K rises as K increases on all datasets, reaching 0.8267, 0.8967, and 0.6105 at K=10 for NYC, TKY, and CA.At K=50, the corresponding values are 0.9422, 0.9806, and 0.8324.Source: D. Effectiveness of Semantic Intent… · p. 9
- ANCHOR outperforms ANCHOR-MLP and ANCHOR-Trans on category-level Acc@1 and achieves the best POI-level Acc@1 across all datasets.The comparison evaluates category prediction accuracy alongside final POI recommendation accuracy.Source: D. Effectiveness of Semantic Intent… · p. 9
- The authors characterize the fine-tuned LLM as extracting semantic intent for subsequent grounding and reranking, rather than serving as a replaceable category classifier.The category estimates use historical visit categories and target-time context.Source: D. Effectiveness of Semantic Intent… · p. 9
E. User Cold-start Analysis
ANCHOR’s advantage varies with users’ history: it generally exceeds GA-LLM for Normal and Very active users, while results for Inactive users depend on the dataset. On NYC, it combines the highest Acc@1 with relatively low prompt-token cost.
E. User Cold-start Analysis
- ANCHOR generally outperforms GA-LLM for Normal and Very active users, with clear gains for Very active users on TKY and CA.The authors suggest richer histories benefit the history-aware reranker’s matching of candidates to user behavior.Source: E. User Cold-start Analysis · pp. 9–10
F. Model Efficiency
- On NYC, ANCHOR achieves the highest Acc@1 with relatively low prompt-token cost, outperforming alternatives on the accuracy–input-cost trade-off.Fig. 5 compares Acc@1 with prompt-token ranges, averages, and medians. The authors attribute efficiency mainly to field-based prompts and compact category-level intent estimation.Source: Figure 5 · p. 10, F. Model Efficiency · p. 10
E. User Cold-start Analysis
- For Inactive users, ANCHOR beats GA-LLM on CA but trails it on NYC and TKY, leaving extreme cold-start users challenging.The authors suggest limited history may restrict the benefit of history-aware reranking.Source: E. User Cold-start Analysis · p. 10
G. Parameter Sensitivity
Increasing candidate coverage does not guarantee better final ranking: both the number of predicted categories and the candidate-pool size involve trade-offs with ranking quality and inference speed. The related-work discussion notes that structure-enhanced methods still mainly predict concrete POIs without explicitly modeling high-level mobility intent.
G. Parameter Sensitivity
- Across three datasets, LLM Category Hit and Candidate Hit generally rise with Kc, while Acc@1 and MRR stabilize or slightly decline at larger values.Fig. 6 examines coverage-oriented and final recommendation metrics across different Kc values.Source: Figure 6 · p. 10, G. Parameter Sensitivity · p. 10
- Larger Kc can add weakly related candidates after plausible categories are covered, weakening category selectivity and burdening the reranker.The analysis recommends balancing uncertainty tolerance with candidate-level discrimination rather than maximizing coverage alone.Source: G. Parameter Sensitivity · p. 10
- As N increases, evaluation speed drops, so larger candidate pools trade improved coverage for higher reranking cost and potential ranking noise.The analysis attributes the cost increase to scoring each candidate against user history and target context.Source: G. Parameter Sensitivity · p. 11
- Candidate Hit rises as N grows, but Acc@1 does not consistently follow; on NYC, Acc@1 peaks at N = 1500 and then declines.Fig. 7 tracks performance and efficiency as N increases from 100 to 3000.Source: Figure 7 · p. 11, G. Parameter Sensitivity · p. 11
A. Next POI Recommendation
- Structure-enhanced methods model POI-level transitions but generally lack explicit modeling of high-level mobility intent.The discussion names trajectory-flow, spatio-temporal hypergraph, and adaptive dependency-graph approaches among these methods.Source: A. Next POI Recommendation · p. 11
B. Language Models for Mobility Data
LLMs bring semantic reasoning to mobility data, but textualizing trajectories or generating POI identifiers does not itself ground predictions in valid, localized entities. ANCHOR bridges category-level intent to concrete POIs using mobility evidence and reranking, and the paper reports consistent gains over diverse baselines while noting a dependence on historical data.
B. Language Models for Mobility Data
- Trajectory-to-language, POI-enrichment, and generative methods apply LLMs to mobility prediction, but textualization or POI-ID generation alone does not ensure valid entity grounding.The cited approaches include trajectory prompting and similarity, zero-shot next-POI recommendation, and generative recommendation using semantic IDs.Source: B. Language Models for Mobility Data · p. 12
C. Semantic-to-Entity Grounding
- For localized POI prediction, category-level intent can map to many candidates whose relevance depends on history, locality, temporal context, and category compatibility.Reliable prediction therefore requires candidate construction and evidence integration over structured mobility records.Source: C. Semantic-to-Entity Grounding · p. 12
VI. CONCLUSION
- ANCHOR estimates category-level intent, grounds it into POI candidates using history and preference evidence, then applies history-aware reranking instead of unconstrained POI generation.The conclusion reports consistent outperformance over diverse next-POI recommendation baselines on three real-world check-in datasets.Source: C. Semantic-to-Entity Grounding · p. 12, §VI · p. 12
- ANCHOR depends on sufficient historical mobility data for grounding and reranking, which may limit robustness for users with very short check-in histories.The paper identifies adaptive grounding for different user contexts and mobility scenarios as future work.Source: §VI · p. 12