PaperlayerView exact revision on arXiv

Source-linked AI summary

SPAR: Enhancing Industrial-Scale Generative POI Recommendation via Real-World Spatial Perception

Fangye Wang, Yunjin Gu, Haowen Lin, Yifang Yuan, Song Yang, Xiaojiang Zhou, Pengjie Wang

arXiv:2609.02062v1cs.IR

TL;DR

Generative POI recommenders often model behavioral sequences without explicitly learning the distance, direction, and reachability relations that govern movement through real cities. SPAR addresses this with spatially grounded identifiers, multi-granular geospatial pre-training, and task-vector anchoring, achieving state-of-the-art performance across public and industrial datasets.

  • Problem

    Existing generative POI recommendation methods leave urban spatial knowledge weakly encoded or absent, despite users physically traveling according to distance, direction, and reachability.

  • Method

    SPAR combines coordinate-aware SI-SID, MG-CPT on 25 multi-tier geospatial datasets, and TV-SFT with a frozen spatial task vector.

  • Results

    SPAR achieves SOTA performance across two public benchmarks and four industrial datasets, with recommendations becoming geographically closer and practically reachable.

  • Takeaways & Limitations

    Grounding the interest space in real urban spatial knowledge supports accurate generative POI recommendation within the reported evaluation settings.

Abstract

from arXiv · show

Generative Point-of-Interest (POI) recommendation, autoregressively generating a target POI's semantic ID (SID), holds great promise for Location-Based Services, where a recommendation helps only if the user can reach it. Yet, existing methods operate within an interest space defined by behavior sequences and collaborative signals, where geography enters only as a textual attribute of the SID, leaving no explicit mechanism to learn or preserve how urban places are related by distance, direction, and reachability; their predictions are thus behaviorally plausible yet far from the user's real-time location. We argue that such services require injecting real urban spatial knowledge into the interest space, rather than inferring geography from behavior alone. Hence, we propose SPAR, a unified framework whose three synergistic stages jointly construct, cultivate, and preserve urban spatial knowledge: (1) at the tokenization level, Spatially-Intrinsic SID (SI-SID) explicitly encodes longitude--latitude coordinates into a sinusoidal geospatial embedding and fuses it with the textual semantic embedding, producing identifiers via RQ-Kmeans that are simultaneously semantically and geographically consistent; (2) at the cognition level, Multi-Granular Geospatial CPT (MG-CPT) continually pre-trains the base LLM on 25 curated geospatial datasets organized into three tiers of basic attributes, pairwise relations, and city-scale navigation, so that scattered POIs cohere into a connected urban space; and (3) at the adaptation level, Task-Vector Anchored SFT (TV-SFT) anchors the acquired spatial knowledge as a frozen parameter-space task vector to prevent its catastrophic forgetting during behavioral fine-tuning, thereby fusing the two spaces. Extensive quantitative and visualization experiments on two public and four industrial-scale datasets demonstrate the effectiveness of SPAR.

1 Introduction

Next-POI recommendation is central to large-scale Location-Based Services, but existing generative methods mainly model behavioral interest space and inadequately encode real urban geography. SPAR addresses this by constructing, cultivating, and preserving spatial knowledge across tokenization, pre-training, and fine-tuning.

  • Location-Based Services use next-POI recommendation to predict where users will go from historical behavior and real-time context.Its quality affects decision-making efficiency and user experience across platforms serving billions of users and POIs.
  • Existing pipelines operate largely in behavioral co-occurrence space, leaving urban spatial knowledge imprecisely encoded in tokens and largely absent from training.
  • Textual or indirect geographic signals produce identifiers dominated by semantic variance, causing the identifier space to lack spatial continuity.Longitude–latitude digits lack metric structure, while collaborative and contrastive signals cannot replace explicit geographic measurement.
  • Behavior-focused training overlooks that users physically travel through cities and consider distance, direction, and reachability, while generic pre-training lacks fine-grained city knowledge.Existing CPT over textual POI attributes does little to learn road-network connectivity.
  • SPAR injects urban spatial knowledge through SI-SID, MG-CPT, and TV-SFT, respectively constructing, cultivating, and preserving it within generative recommendation.The framework combines coordinate-aware identifiers, multi-tier geospatial pre-training, and task-vector anchoring during behavioral adaptation.

2 Related Works

Next-POI recommendation has progressed from sequential models to LLM-based generative recommendation, but existing generative approaches remain insufficiently grounded in physical urban space. SPAR distinguishes itself by making geography central throughout the recommendation pipeline.

  • Traditional next-POI methods evolved from sequential models using Markov chains, matrix factorization, RNNs, and Transformers to capture behavioral preferences.
  • LLM-based generative recommendation directly generates item identifiers, replacing cascaded retrieval-and-ranking pipelines with autoregressive generation.TIGER introduced hierarchical SIDs, while later systems adopted RQ-Kmeans and industrial end-to-end frameworks.
  • Existing generative recommendation models overlook that next-POI predictions must correspond to real physical space rather than behavior and interest co-occurrence alone.
  • SPAR injects real urban spatial knowledge into every stage to ground behavioral fitting in city geography.

3 Problem Formulation

The task models users, POIs, interaction trajectories, and real-time conditions, then predicts the next POI from the user’s history and context. SPAR performs this prediction generatively by autoregressively decoding the next POI’s SID.

  • The formulation defines U as M users and P as N POIs, with each POI represented by coordinates, address, contextual attributes, and an RQ-Kmeans-generated SID.
  • Each user trajectory records interaction events in reverse chronological order.
  • An interaction records the user, POI, timestamp, condition, and action, including navigation, reservation, booking, collection, query, and current coordinates in AMAP.
  • Next-POI recommendation predicts POI p_n+1 by maximizing its conditional probability given the user, trajectory, and real-time condition.
  • Unlike discriminative methods that score all candidates, SPAR autoregressively decodes the next POI’s SID from the historical SID sequence conditioned on real-time context.

4 Methodology: SPAR

SPAR builds urban spatial knowledge into generative POI recommendation through spatially intrinsic identifiers, multi-granular geospatial pre-training, and task-vector-anchored fine-tuning. Together, these stages fuse semantic and geographic information, teach spatial relations, and preserve them during behavioral adaptation.

  • SPAR combines SI-SID, MG-CPT, and TV-SFT to construct, cultivate, and preserve urban spatial knowledge for generative POI recommendation.SI-SID builds the identifier space, MG-CPT develops spatial reasoning, and TV-SFT anchors spatial knowledge during adaptation.
  • SI-SID: SI-SID encodes normalized longitude–latitude coordinates with a sinusoidal encoder, fuses the resulting geospatial embedding with semantic features, and quantizes them using residual K-means.The encoder uses multiple frequency bands, while L2 normalization prevents either semantic or geographic signal from dominating clustering.
  • SI-SID: Sinusoidal encoding makes embedding similarity depend on coordinate displacement, and its continuity encourages nearby POIs to receive adjacent embeddings and share SI-SID prefixes.These properties make geographic proximity an intrinsic tendency of the identifier space after residual quantization.
  • SI-SID: The fused representation is quantized layer by layer with RQ-Kmeans into four-level SI-SIDs, whose fourth level disambiguates POIs that collide on earlier codewords.The first three levels are learned codebooks; the final index guarantees identifier uniqueness.
  • MG-CPT: MG-CPT continually pre-trains the base LLM on 25 datasets covering POI and road attributes, pairwise spatial relations, and city-scale navigation.The datasets jointly characterize entities, surrounding road relationships, and dynamic road-system behavior across millions of training instances.
  • TV-SFT: TV-SFT separates behavioral adaptation from spatial preservation by training the base model for user behavior while freezing the MG-CPT task vector and applying a trainable low-rank LoRA adapter.The task vector represents the parameter change induced by spatial pre-training, and only WBase and ΔLoRA remain trainable during this stage.

5 Experiments

Across public and industrial-scale evaluations, SPAR consistently improves generative POI recommendation by constructing, learning, and preserving explicit urban spatial knowledge. Its gains extend to proximity, cold-start robustness, hierarchical spatial organization, and each ablated training stage.

  • 5.2.1 Results on Public Benchmarks: SPAR achieves state-of-the-art performance across all metrics on the NYC and TKY public benchmarks.
  • 5.2.1 Results on Public Benchmarks: SPAR-8B surpasses PLUM (8B) across all public-benchmark metrics, with relative improvements of 11.91%–19.29%, while SPAR-4B outperforms every baseline.The result is attributed to spatial reinforcement across SI-SID, MG-CPT, and TV-SFT.
  • 5.2.2 Results on Industrial Datasets: On four industrial datasets, SPAR-8B gains 38.32% on average over PLUM across R@5/N@5/M@5, while SPAR-0.6B surpasses every 8B baseline on all twelve metrics.Average gains for SPAR-4B and SPAR-0.6B are 28.54% and 9.92%, respectively.
  • 5.3 Ablation Study: SI-SID and MG-CPT contribute sequential average improvements of 10.76% and 15.38%, indicating complementary gains from geographic identifier construction and spatial-relation learning.The ablation compares TextOnlyGR → SPAR-NoCPT → SPAR-FullSFT.
  • 5.3 Ablation Study: TV-SFT adds relative improvements of 2.15%–17.90% across four datasets by freezing the MG-CPT task vector while adapting behavioral patterns with LoRA.Tianjin shows the largest gains: 15.32% R@5, 17.05% N@5, and 17.90% MRR@5.
  • 5.4 Cold-Start Performance: SPAR improves cold-start Recall@5 over TextOnlyGR by 47.6% on Shanghai and 48.0% on Zhejiang, narrowing the cold-to-warm gap to 18.5%–20.6%.The gains are larger than in warm-start settings, where improvements are 30.7% and 14.9%.

6 Conclusion

The conclusion argues that generative POI recommendation should ground behavioral interests in real urban space. SPAR does so through three synergistic stages, and experiments support its effectiveness for accurate, reachable recommendations.

  • Conclusion: SPAR injects real urban spatial knowledge into generative POI recommendation through three synergistic stages.SI-SID constructs the geographic foundation, while MG-CPT and TV-SFT cultivate and preserve spatial knowledge.
  • Conclusion: Experiments on two public benchmarks and four industrial datasets show that SPAR achieves state-of-the-art performance.The evaluation combines quantitative and qualitative evidence across six datasets.
  • Conclusion: The experiments also indicate that SPAR recommendations become geographically closer and practically reachable.The conclusion attributes this outcome to grounding interest modeling in real urban spatial knowledge.

A Theoretical Analysis of the Geospatial Encoder

The theoretical analysis explains why the sinusoidal geospatial encoder can preserve geographic topology. It establishes exact properties of the Fourier features and examines their survival through the learned projection.

  • Theoretical motivation: The appendix analyzes why the sinusoidal geospatial encoder produces embeddings whose geometry mirrors geographic topology.This property is important because SI-SID inherits it through residual quantization.
  • Analysis plan: The analysis first states two exact properties of the Fourier features used by the encoder.The subsequent discussion verifies whether these properties survive the learned projection.
  • Analysis plan: The learned projection is empirically examined to determine whether the encoder’s theoretical spatial structure remains present in final embeddings.The appendix connects the Fourier-feature analysis to the geospatial embeddings used by SI-SID.

A.1 Exact Properties

The geospatial encoder is designed to make similarity depend on coordinate displacement while keeping nearby coordinates close in embedding space. Continuity is preserved by the projection, whereas monotonicity after projection is only qualitative and empirical.

  • Exact properties: The shift-invariant kernel depends on coordinate differences rather than absolute positions, making embedding similarity decrease with geographic displacement.The stated monotonicity covers intra-city displacements within the analyzed range.
  • Exact properties: Lipschitz continuity guarantees that nearby coordinates produce nearby Fourier features, and the MLP preserves this continuity before RQ-Kmeans.This provides the prerequisite for adjacent POIs to receive adjacent SI-SIDs.
  • Exact properties: The encoder uses exponentially spaced frequencies so different bands respond to coordinate differences at different rates.The MLP learns a weighted combination of these frequency bands during end-to-end training.
  • Qualification: The exact continuity property survives the MLP, while distance–similarity monotonicity after projection is preserved qualitatively and verified empirically.The distinction limits the strength of the post-projection claim.
  • Scope: Within a single city, normalized-coordinate differences are treated as approximately proportional to geodesic distance.This assumption supports transferring the monotonicity argument to real distances at next-POI recommendation scale.

A.2 Empirical Verification

Experiments on a 30 × 30 Beijing coordinate grid show that the projected embeddings preserve grid regularity and correlate with geographic distance. This encoder-level geometry is consistent with the spatial and semantic hierarchy observed in SI-SID construction.

  • Experimental setup: A 30 × 30 Beijing coordinate grid provides 900 locations for evaluating the final geospatial embeddings after MLP projection.The grid covers roughly 44 km × 33 km of an urban region.
  • Grid topology: The t-SNE projection forms an evenly spaced grid that mirrors the original coordinates, indicating position-independent embedding changes.No region of the city is reported as stretched or compressed in the projection.
  • Metric preservation: Pearson r=0.893 shows that embedding distance grows monotonically and nearlinearly with geographic distance.Nearby coordinates receive nearby embeddings, while farther locations are mapped farther apart.
  • Connection to SI-SID: The verified encoder geometry supports spatially continuous SI-SIDs, with first-level clusters organized mainly by geographic proximity.Within a first-level cluster, second-level sub-clusters separate geographically overlapping POIs by semantic category.
  • Connection to MG-CPT: MG-CPT complements this representation by training on 25 geospatial datasets so POIs become interconnected places rather than isolated tokens.Its corpus spans points, lines, and flows over urban space.

B.1 Three Tiers of Datasets

MG-CPT organizes urban spatial knowledge into three tiers, combining object attributes, pairwise relations, and city-scale movement information. Together, these datasets connect POIs, roads, and travel paths into an internalized urban space.

  • Basic attributes: The basic-attribute tier uses 13 datasets to map POI SI-SIDs, coordinates, and attributes, while describing individual roads.It includes twelve POI datasets and one road dataset covering class, length, lanes, and speed limit.
  • Static road network: Road-feature tasks encode the static skeleton of the city through road class, orientation, and connectivity.These features describe arterials, ring roads, and expressways.
  • Dynamic movement flow: Navigation data encode dynamic movement as origin POI, traversed road sequence, and destination POI connections.Aggregated routes connect isolated POIs through real travel paths.
  • Pairwise relations: Distance and direction tasks use geographic measurements, including planar distance and eight-way directional classification.The planar approximation uses 85 kilometers per longitude degree and 111 kilometers per latitude degree at the city’s latitude.
  • Task templates: The dataset templates cover POI coordinate, reverse-coordinate, category, nearby-POI, road-description, distance, and direction queries.Templates include both multiple-choice and open-ended formats, with coordinate-based and SID-based variants.

B.5.3 Relational: point to point (9 tasks).

The relational tier models point-to-point geography through nearby-POI, distance, direction, and route tasks. These tasks range from coordinate-based geometric reasoning to SID-based recall of spatial relations and navigation information.

  • Nearby POI: Nearby-POI tasks select the closest coordinate among candidate locations by comparing distances to the reference point.The template is multiple-choice and requires outputting the correct option letter.
  • Distance: Distance tasks judge straight-line separation from coordinates or from POI SIDs that encode the needed spatial information.The corpus distinguishes coordinate-based and no-coordinate variants.
  • Direction: Direction tasks classify one POI relative to another using coordinate differences or recalled coordinates from their SIDs.The open-ended format requires the answer first, followed by component-wise reasoning.
  • Navigation: Route-distance tasks estimate road-network distance and travel time between origin and destination POI SIDs.The template explicitly distinguishes road-network distance from straight-line distance.

C Details of the Spatial Cognition Evaluation Benchmark

The spatial cognition benchmark evaluates MG-CPT with automatically generated, city-specific questions spanning the same three-tier organization as its training corpus. It uses disjoint data, standardized prompts, and both multiple-choice and open-ended answer formats.

  • Benchmark construction: Each industrial dataset receives an 18-task benchmark with 2,000 questions per task, generated from city POI, road, and navigation data.The evaluation questions are strictly disjoint from the MG-CPT training corpus.
  • Task organization: The benchmark follows the three-tier organization of basic attributes, relational tasks, and systemic city-scale knowledge.This mirrors the structure used to organize MG-CPT’s spatial cognition training data.
  • Evaluation protocol: Each evaluation item includes a task-specific system prompt defining a city-geography expert role and the expected output format.The final input concatenates the system prompt with the question.
  • Answer formats: Multiple-choice tasks require the model to output an option letter, whereas open-ended tasks request answers such as direction, distance, or district.The two formats use distinct standardized prompts.
Loading 2609.02062v1…