Where to Move Next: Zero-shot Generalization of LLMs for Next POI Recommendation

📝 Paper Summary

Recommender Systems Spatial-Temporal Modeling

LLMmove is a zero-shot prompting framework that enables generic Large Language Models to recommend the next Point-of-Interest by explicitly incorporating user preferences, pre-calculated geographical distances, and sequential transition patterns.

Core Problem

Traditional POI recommendation requires training expensive task-specific models on large datasets, while off-the-shelf LLMs struggle with spatial reasoning (distances) and specific mobility sequences.

Why it matters:

Training task-specific models for every city/region consumes extensive computational resources
Standard LLMs hallucinate or ignore geographical constraints (Tobler's First Law), recommending popular but physically distant locations
Existing zero-shot LLM recommenders focus on e-commerce/movies and fail to capture the spatial and sequential dependencies unique to human mobility

Concrete Example: A user visits a 'Gym' then a 'Neighborhood' area. A standard LLM might suggest a popular 'Museum' far away. LLMmove identifies the 'Gym -> Neighborhood' pattern and the user's preference for short trips, recommending a nearby 'Subway' or 'Home' instead.

Key Novelty

LLMmove (Prompting Framework)

Decomposes mobility reasoning into four explicit prompt components: long-term preference, recent preference, geographical distance, and sequential transitions
Bypasses LLM calculation errors by pre-calculating spatial distances for candidates and feeding them as textual context (e.g., 'Distance: 0.5km')
Uses a distance-based ordering strategy for candidate presentation to mitigate LLM sensitivity to input order

Evaluation Highlights

Outperforms LLMMob(+Geo) variant, demonstrating that explicit prompting for spatial distance and transition patterns is superior to embedding-based approaches in zero-shot settings
Significantly outperforms standard zero-shot sequential baselines (CZSR, LLMRank) which lack geospatial consideration
Ordering candidate POIs by ascending distance (closest first) yields remarkably better performance than random or descending order

Breakthrough Assessment

7/10

Novel application of LLMs to spatial mobility without training. Smartly circumvents LLM math limitations via pre-calculation. Lack of numeric results in the provided text limits verification of magnitude.

⚙️ Technical Details

Problem Definition

Setting: Next POI recommendation given a user's check-in trajectory

Inputs: User history check-ins (long-term), recent trajectory, candidate POI set

Outputs: Top-K predicted POIs for the next visit

Pipeline Flow

Data Preprocessing (History Partitioning)
Distance Calculation (External Tool)
Prompt Construction
LLM Inference (Generation & Ranking)

System Modules

Data Partitioner (Input Processing)

Split user history into long-term check-ins (historical preference) and recent check-ins (current context)

Model or implementation: Rule-based script

Distance Calculator (Input Processing)

Compute exact geographical distances between the user's last location and all candidate POIs (overcoming LLM math failure)

Model or implementation: Deterministic calculation (Haversine formula implied)

LLM Inference

Reason over the four factors (Long-term, Recent, Geo, Seq) to rank candidates and provide explanations

Model or implementation: gpt-3.5-turbo

Novel Architectural Elements

Integration of hard-coded distance calculations into the prompt context to ground the LLM's spatial reasoning
Use of 'Distance-ascending' candidate ordering specifically to bias the LLM towards Tobler's First Law

Modeling

Base Model: gpt-3.5-turbo

Training Method: Zero-shot prompting only (no fine-tuning)

Training Data:

Datasets: NYC and TKY (Foursquare check-ins, Apr 2012 - Feb 2013)
Split: 80% training (used only for long-term history context), 10% validation, 10% test
Candidate set: Ground truth + 100 randomly sampled POIs

Compute: Not reported in the paper

Comparison to Prior Work

vs. LLMMob: LLMmove focuses on spatial distance and sequence transitions rather than just temporal information; LLMmove allows recommending new locations vs. only historical ones
vs. CZSR/LLMRank: LLMmove explicitly integrates geospatial distance (calculated externally) into the prompt, whereas general sequential methods ignore spatial laws
vs. Dist (Nearest Neighbor): LLMmove combines distance with semantic user preferences, whereas Dist ignores user interest [not cited in paper, but implied simple baseline]

Limitations

Sensitive to the order of candidate POIs (performance drops if not sorted by distance)
Cannot accurately calculate spatial distances internally; relies on external calculation fed as input
Zero-shot performance gap compared to fully supervised state-of-the-art models (implied by the nature of zero-shot vs supervised)

Reproducibility

Code: https://github.com/LLMMove/LLMMove

Code and datasets are publicly available at https://github.com/LLMMove/LLMMove. The prompt structure is described conceptually (incorporating 4 factors) but the exact full prompt text is not in the paper body.

📊 Experiments & Results

Evaluation Setup

Next POI prediction on real-world check-in datasets

Benchmarks:

NYC Dataset (Next POI Recommendation)
TKY Dataset (Next POI Recommendation)

Metrics:

Acc@K (Top-K Accuracy)
MRR (Mean Reciprocal Rank)
Statistical methodology: Not explicitly reported in the paper

Main Takeaways

Geographical distance is the single most critical factor; the 'Dist' (nearest neighbor) baseline alone outperforms complex generic sequential recommenders.
LLMmove achieves the best performance among zero-shot methods by combining distance with user preference, outperforming the LLMMob variants.
Ablation studies reveal that 'Long-term preference' and 'Geographical influence' are the most impactful components of the prompt.
Candidate ordering significantly impacts LLM output: sorting candidates by ascending distance (closest first) yields much better results than random or frequency-based sorting.

📚 Prerequisite Knowledge

Prerequisites

Basics of Recommender Systems (Sequential Recommendation)
Understanding of Large Language Models (Prompting)
Spatial analysis concepts (Coordinates, Distance)

Key Terms

POI: Point of Interest—a specific geographical location users visit (e.g., a coffee shop, gym)

Zero-shot: Using a pre-trained model to perform a task without any gradient-based fine-tuning on task-specific data

Check-in: A record of a user visiting a specific POI at a specific time

LBSN: Location-Based Social Networks—platforms like Foursquare where users share locations

Tobler's First Law: A geography principle stating 'Everything is related to everything else, but near things are more related than distant things'

Acc@K: Top-K Accuracy—metric checking if the ground truth is within the top K recommendations

MRR: Mean Reciprocal Rank—metric evaluating the ranking position of the correct answer