PaperlayerView exact revision on arXiv

Source-linked AI summary

SemPOI-RL: Aligning LLM Semantic Reasoning for Interpretable Out-of-Town POI Sequential Generation

Yunqi Liu, Yang Zhang, Ruixing Zhang, Liangzhe Han, Yi Qiao, Tongyu Zhu, Leilei Sun

arXiv:2608.30399v1cs.CLcs.AI

TL;DR

OOT POI generation must connect transferable travel intent and semantic reasoning with coherent, structurally constrained destination sequences, a gap left by latent ID-based and direct-LLM approaches. SemPOI-RL uses LLM-inferred travel styles, position-aware SPAM generation, and reinforcement learning, outperforming traditional and LLM baselines while providing phase-specific style attribution. Its scope remains bounded by latent prototypes, pseudo-label dependence, known benchmark constraints, catalog exposure, pipeline cost, and low exact-POI HR.

  • Problem

    OOT recommendation lacks an interpretable trainable interface that grounds high-level travel-style inference into position-aware sequence generation under cross-city interest drift.

  • Method

    SemPOI-RL fine-tunes an LLM for destination travel styles, grounds them through SPAM in a masked autoencoder, and optimizes style generation with trajectory rewards.

  • Results

    SemPOI-RL consistently outperforms traditional sequential and LLM-based baselines while providing interpretable style attribution across trip phases.

  • Takeaways & Limitations

    The framework provides a more transparent link between semantic reasoning and structured cross-city trajectory recommendation.

  • Takeaways & Limitations

    The method depends on latent prototypes and LLM-generated pseudo-labels, assumes known start/end points, lengths, and destination catalogs, and has higher cost with low exact-POI HR.

Abstract

from arXiv · show

Large language models (LLMs) exhibit strong semantic reasoning and open-ended generation abilities, but aligning these abilities with structured sequential generation remains challenging. This challenge is particularly evident in out-of-town (OOT) POI sequence generation, where a model must infer transferable travel intent from a user's hometown behaviors, adapt to cross-city interest drift, and generate a coherent destination trajectory under structural constraints. Existing approaches either rely on latent ID-based transfer with limited interpretability or directly use LLMs for sequence generation without explicitly grounding inferred semantics into position-aware predictions. To address this gap, we propose SemPOI-RL, a framework that aligns LLM semantic reasoning with structured sequence generation for interpretable OOT recommendation. Specifically, we first fine-tune an LLM to infer destination-oriented travel styles from users' hometown trajectories, using natural language as an interpretable semantic intermediate. We then introduce a Semantic POI Alignment Module (SPAM) to ground these inferred styles into a style-conditioned masked autoencoder for position-aware trajectory generation. Finally, we apply reinforcement learning with recommendation-oriented rewards to align LLM-generated styles with downstream sequence quality. Experiments on two real-world datasets show that SemPOI-RL consistently outperforms both traditional recommenders and direct LLM baselines, while providing interpretable style attribution across different phases of a trip. The code is available at https://github.com/Wind-Flipped/SemPOI-RL .

1 Introduction

OOT POI generation must transfer travel intent across cities while producing ordered, coherent trajectories under interest drift and structural constraints. SemPOI-RL aligns interpretable LLM-inferred styles with position-aware sequence generation and recommendation rewards.

  • OOT recommendation requires transferable travel intent and coherent ordered destination trajectories, not merely next-POI prediction.
  • Existing latent ID-based and direct-LLM approaches offer limited interpretability and do not align textual travel styles with position-aware trajectory prediction.
  • SemPOI-RL treats natural-language travel style as an intermediate representation trained from destination behavior, grounded at each trajectory position, and optimized by the resulting sequence.
  • GRPO uses hit rate, recall, category consistency, and diversity rewards to refine LLM-generated styles for downstream trajectory quality.
  • Experiments on two real-world datasets show consistent gains over traditional and direct-LLM baselines in accuracy, coherence, and interpretability.
  • SPAM decomposes global styles into temporally varying prototypes and integrates them with a masked autoencoder to dynamically weight style influence across sequence positions.

2 Related Work

Related work progresses from Markovian, spatio-temporal, and graph-based recommenders toward LLM-enhanced preference modeling and explanation. Existing LLM approaches include preference encoding, semantic POI representations, and trajectory generation.

  • Early POI sequence methods modeled transitions probabilistically, while later recurrent and transformer models incorporated temporal intervals and geographic distances.
  • LLMs have entered recommender systems through natural-language understanding, few-shot adaptability, and human-readable recommendation rationales.
  • Prior LLM-based mobility methods encode user preferences, infer occupations or travel motivations, enrich POI representations, and support temporal or next-point prediction.

3 Preliminary

The preliminary formulation defines POIs, check-ins, hometowns, and out-of-town behavior before specifying the cross-city sequence recommendation task. The benchmark uses an unseen user but a destination catalog observed during training.

  • A POI is a spatial item associated with a geographical location, represented with region, longitude-latitude location, and category attributes.
  • A check-in records a user's direct real-world interaction with a POI as the tuple c = (u, t, v).
  • The user's hometown is identified as a region where the user lives or works for a period, using time-weighted recent check-in frequencies.
  • Out-of-town travel behavior combines a user's hometown and destination-region check-in records into the tuple ξ = (u, ch, co, rh, ro).
  • The task learns a recommender from historical records for a new user and destination region, generating an ordered destination POI sequence.
  • The benchmark treats the user as unseen but assumes the destination region and its POI catalog appear in training data, rather than evaluating a completely new city catalog.

4 Methodology

SemPOI-RL bridges destination-oriented travel-style inference from hometown trajectories with position-aware destination trajectory generation. Its three stages use supervised fine-tuning, SPAM with a masked autoencoder, and reinforcement learning to align semantic styles with recommendation quality.

  • 4 Methodology: SemPOI-RL decomposes cross-city recommendation into destination-style inference, semantic grounding in trajectory generation, and reinforcement alignment.The framework treats natural-language travel style as an intermediate representation between LLM inference and structured POI sequence generation.
  • 4 Methodology: The LLM is fine-tuned to predict destination-style summaries from users’ hometown trajectories, using observed destination summaries as supervision.After fine-tuning, it generates a destination-oriented style summary and its embedding.
  • 4 Methodology: SPAM decomposes a global style embedding into multiple prototypical semantic components that can be selectively activated across trajectory positions.This avoids explaining an entire trip with one semantic vector and supports temporal variation across trip phases.
  • 4 Methodology: A diversity loss encourages distinct prototypes, while negative normalized entropy discourages one prototype from dominating every position without forcing uniform assignments.Their combination preserves distinct semantic components and trip-phase specificity.
  • 4 Methodology: The style-conditioned masked autoencoder reconstructs masked intermediate destination POIs from partially observed trajectories while exposing position-wise style assignments.Hometown points and destination endpoints remain visible during training, while a proportion of intermediate destination points is masked.
  • 4 Methodology: GRPO samples multiple style descriptions and updates the LLM using trajectory-derived rewards for hit rate, recall, category consistency, and diversity.The reward is computed from predicted trajectories against ground truth, propagating recommendation quality back to the natural-language style representation.

5 Experiment

SemPOI-RL is evaluated on Foursquare and Yelp against traditional and LLM-based baselines under full out-of-town sequence generation. It achieves stronger overall performance and ablations show that supervised style adaptation, positional semantic alignment, and reward optimization each contribute to the results.

  • 5.1 Experiment Setup: Experiments use Foursquare and Yelp, six metrics, and a setting where start and end POIs plus trajectory length are given while intermediate POIs are generated.The test input excludes other destination check-ins, and all methods receive the same destination catalog.
  • 5.2 Overall Result: SemPOI-RL achieves the best overall performance across both datasets, improving Hit Rate by about 15% on Foursquare and 10% on Yelp over the strongest non-LLM baselines.The comparison includes seven traditional trip recommendation methods and three LLM-related methods.
  • 5.2 Overall Result: SemPOI-RL leads on category and structural measures, with stronger Category Match and lower Edit Distance and Dynamic Time Warping Distance than the baselines.These measures provide complementary evidence when exact POI IDs are difficult to match at precise sequence positions.
  • 5.3 Ablation Study: All ablated variants underperform the full model across most metrics, supporting the necessity of SPAM, supervised fine-tuning, and reinforcement learning.The ablations remove SPAM, skip SFT, or remove RL from the pipeline.
  • 5.3 Ablation Study: Removing supervised fine-tuning causes the largest drop, while removing SPAM or reinforcement learning also degrades performance across trajectory, semantic, and downstream generation measures.SPAM supports temporally varying prototype styles, and reward optimization further refines the LLM-inferred styles for sequence generation.
  • 5.4 Interpretability Analysis: The case study shows a transition from leisure, dining, and entertainment venues to rest and trip-closure POIs across the generated trajectory.Prototype assignments attribute the entertainment-oriented early positions primarily to prototype 3, while later positions involve prototype 1.

6 Conclusion

SemPOI-RL aligns LLM-inferred destination travel styles with position-aware POI sequence generation through semantic decomposition and reinforcement learning. Across two real-world datasets, it outperforms traditional and LLM-based baselines while providing interpretable style attribution across trip phases.

  • 6 Conclusion: SemPOI-RL bridges high-level preference semantics and structured destination trajectory generation for cross-city POI recommendation.It uses an LLM to infer destination travel styles from hometown trajectories as an interpretable abstraction of cross-region interest drift.
  • 6 Conclusion: SPAM decomposes global style signals into temporally varying prototypes and injects them into a style-conditioned masked autoencoder for position-aware POI prediction.This grounds semantic reasoning in different phases of the generated trajectory.
  • 6 Conclusion: Experiments on two real-world datasets show that SemPOI-RL consistently outperforms traditional sequential recommendation models and recent LLM-based baselines.The framework also improves trajectory accuracy and structural coherence while exposing interpretable style attribution across trip phases.

Limitations

SemPOI-RL is constrained by latent prototype interpretability, pseudo-label dependence, benchmark assumptions, fixed catalogs, and deployment costs. Open-ended unseen-city transfer and stronger reliability evidence remain future needs.

  • Prototype styles remain latent and are interpreted indirectly through activated POI categories, while style quality depends on LLM-generated pseudo-labels.
  • The benchmark assumes known start and end POIs plus trajectory length, and trains on fixed destination catalogs rather than testing zero-shot transfer to unseen cities.
  • Open-ended generation and catalog transfer remain future work.
  • The multi-stage pipeline costs more than simple recommenders, exact-POI HR remains low, and deployment requires further reliability evaluation and user studies.

Ethical Considerations

The paper treats POI trajectories as sensitive data and describes anonymization and data minimization safeguards. It also recognizes residual privacy, bias, and stereotype risks in deployment.

  • POI trajectories can reveal users’ routines and private attributes, so user identifiers were anonymized before analysis.
  • The study uses structured check-in records without user-authored free text, does not recover identities, and retains only research-needed fields.
  • Deployment should apply data minimization and access controls because check-ins and generated styles may carry privacy and bias risks.
  • Popularity, regional, and LLM-derived stereotype biases may produce uneven recommendation quality across users and locations.

A Datasets

Experiments use Foursquare and Yelp cross-region travel trajectories constructed from users with hometown and out-of-town check-ins. Destination data supervises training and evaluation but is not a test-time input.

  • Foursquare and Yelp provide users with check-in activity in both hometown and out-of-town locations for cross-region trajectory construction.
  • Users with fewer than three destination check-ins or trips outside one hour to 30 days are removed, as are POIs visited fewer than twice.
  • At test time, the model receives only the hometown trajectory, destination start/end POIs, and trajectory length.
  • The destination trajectory is used solely to construct Stage 1 SFT pseudo-labels and compute training rewards and evaluation metrics.
  • Table 3 distinguishes hometown trajectory H, destination trajectory D, and query Q containing destination start/end POIs plus length.

B Metrics

Evaluation measures exact positional accuracy, POI coverage, sequence alignment, category and region consistency, and diversity-related behavior across two datasets and multiple baselines. The study also reports sensitivity, ablation, and reward-weight analyses.

  • Metrics: Hit Rate measures whether each predicted POI exactly matches the ground-truth POI at the same position.
  • Metrics: Recall Rate measures the fraction of unique ground-truth POIs covered regardless of prediction positions.
  • Metrics: Edit Distance uses normalized Levenshtein distance, while Dynamic Time Warping measures alignment cost from 0/1 POI matching along a warping path.
  • Analysis: The analysis covers two datasets, prototype count M, diversity regularization λD, ablations, and RL reward-weight allocations.
  • Metrics: Category Match and Region Match measure position-wise consistency in POI categories and regions, respectively.
  • Baselines: Experiments compare SemPOI-RL with recurrent, transformer, graph, hypergraph, ODE-based, search-based, and LLM-based recommendation approaches.

E Hyper-parameter Analysis

The analysis examines hyper-parameter sensitivity, reward weighting, statistical reliability, and computational cost. SemPOI-RL remains robust under moderate changes, while its preferred settings balance representational coverage, alignment quality, and recommendation performance.

  • Prototype styles: M = 8 and λD = 0.1 provide the best overall balance across datasets.Larger prototype pools can over-fragment the style space, while smaller pools cannot adequately represent different temporal trip phases.
  • Prototype styles: Prototype count trades off temporal coverage against fragmentation and generalization.On Foursquare, excessively fine-grained prototypes reduce both HR and RR; with M = 16, the top-8 prototypes receive about 90% of attention mass.
  • Diversity regularization: λD balances prototype diversity: very small values produce homogeneous styles, whereas overly large values hurt recommendation accuracy.The coefficient controls diversity among prototype style embeddings and affects stage-specific semantic capture.
  • RL reward: The reward weights (2, 0.5, 0.5, 1) prioritize Hit, while Recall, Category, and Diversity encode complementary trajectory objectives.The design emphasizes correctly positioned POI hits, semantic consistency, and non-repetitive trajectories.
  • Statistical reliability: SemPOI-RL is consistently and significantly better than SPOT-Trip on Hit Rate, Recall Rate, Edit Distance, and DTW across both datasets.The comparison uses 3 independent random seeds, mean±std reporting, and paired t-tests with p < 0.05; variance across seeds remains small.
  • Efficiency: The full test set finishes within 1 hour on four NVIDIA A100 GPUs, with inference dominated by style generation and one MAE forward pass.Direct LLM baselines instead use a simulated pre-ranked set of 100 destination candidates because prompting tens of thousands of POIs is infeasible.
  • Efficiency: SemPOI-RL is slower than simple recommenders, but its measured cost is described as moderate relative to its accuracy and interpretability gains.The MAE scores the destination catalog directly, unlike the direct LLM baseline protocol that assumes a strong coarse-ranking stage.

F.3 Quantitative Interpretability Analysis

The quantitative analyses evaluate style fidelity, prototype semantics, and user-level changes before and after reinforcement learning. They show that SFT and RL improve alignment between inferred destination styles and generated trajectories while exposing prototype-to-category correspondences.

  • Style fidelity via text similarity: Style cosine similarity increases monotonically from the pre-SFT model to SFT and then to the RL-tuned model on both datasets.Similarity is measured between the inferred destination style and the ground-truth destination style summary using Qwen3-Embedding-4B.
  • Prototype-to-category statistics: Prototype semantics are characterized through the distributions of 12 coarse-grained POI categories activated by each latent prototype.The analysis compares activations separately for ground-truth and predicted trajectories; prototypes are not directly converted into readable text.
  • Statistical reliability: The multi-seed comparison reports mean±std across 3 seeds and marks statistically significant differences using paired t-tests with p < 0.05.The table compares SemPOI-RL with the strongest baseline, with bold indicating the better mean.
  • Efficiency analysis: The efficiency analysis reports wall-clock cost on four NVIDIA A100 GPUs.The computational-cost results are presented as a separate table-level analysis.
  • Multi-sample comparison before and after RL: For uid=781, RL raises style similarity from 0.654 to 0.727 and trajectory Hit from 0.143 to 0.286, with Recall 0.400.RL refocuses the inferred style on destination nightlife and dining after SFT wording drifts toward hometown routines.
  • Multi-sample comparison before and after RL: For uid=27, RL follows SFT in aligning the style toward a cultural and social characterization, with the case reaching Hit 0.333 and Recall 0.667.The table caption reports the final case-level Hit and Recall values.
  • Stage-1 prompt design: Stage 1 uses separate prompts to summarize destination styles and infer likely destination styles from hometown trajectories.The prompts request concise natural-language, style-level descriptions rather than POI lists, with examples spanning leisure, nightlife, food, culture, business, outdoors, and shopping.
Loading 2608.30399v1…