MRP-LLM: Multitask Reflective Large Language Models for Privacy-Preserving Next POI Recommendation

📝 Paper Summary

Next Point-of-Interest (POI) Recommendation Privacy-Preserving Recommendation LLM for Recommendation

MRP-LLM improves next Point-of-Interest recommendation by using LLMs to reflectively extract fine-grained preferences and summarize neighbor patterns, while applying differential privacy to protect sensitive user data.

Core Problem

Existing LLM-based POI recommenders rely on zero-shot prompting that fails to extract fine-grained spatiotemporal preferences or leverage collaborative signals (patterns from similar users), while directly exposing sensitive check-in history to the model.

Why it matters:

Directly sending raw check-in sequences to cloud-based LLMs risks leaking sensitive personal information like home addresses and daily habits
Standard LLM prompting lacks the 'collaborative filtering' capability of traditional recommenders, leading to lower accuracy in sparse data scenarios
Current approaches treat user history as a flat text sequence, missing complex transition rules (e.g., 'gym after work') and temporal constraints

Concrete Example: A user visits a 'Department Store' at 6 PM. A standard zero-shot LLM might recommend the nearest popular spot. MRP-LLM reflects on history to realize the user has a 'shopping -> dining' transition preference and retrieves neighbor data showing similar users favor a specific restaurant region at that hour, correctly recommending a restaurant.

Key Novelty

Multitask Reflective Preference Extraction with Neighbor Retrieval

Decomposes recommendation into subtasks (category, region, distance prediction) using Chain-of-Thought (CoT) and self-reflection on historical segments to update a structured user preference knowledge base
Injects collaborative signals by identifying geographical, semantic, and social neighbors, then using the LLM to summarize their preferences as context for the target user's recommendation
Protects privacy by perturbing inputs (OUE for sequences, noise for distributions, geo-obfuscation for coordinates) before they ever reach the recommender system

Architecture

The complete workflow of MRP-LLM, from preference extraction to final recommendation

Evaluation Highlights

Outperforms LLM-based baseline LLMMove by 16.8% in Accuracy@1 on the Phoenix dataset without privacy constraints
Achieves comparable performance to SOTA conventional models (STAN) on the Singapore dataset (0.6000 vs 0.6000 Acc@1)
Maintains high utility under privacy protection, with only a 1.3% drop in Accuracy@1 compared to the non-private version on the Singapore dataset

Breakthrough Assessment

7/10

Strong integration of collaborative signals into LLM inference and a rigorous privacy framework. While not fine-tuning the model, the reflective architecture significantly boosts zero-shot utility.

⚙️ Technical Details

Problem Definition

Setting: Next Point-of-Interest (POI) recommendation given historical check-ins, social relations, and current context

Inputs: User history sequences (POI, category, region, distance), current sequence context, and social relationship matrix S

Outputs: Ranked list of candidate POIs for the next visit l_{k+1}

Pipeline Flow

Data Perturbation: Input Processing (Local Device) -> Privacy Transmission
Preference Extraction: History Segments -> LLM Reflection -> Knowledge Base
Neighbor Retrieval: Similarity Calculation -> Neighbor Selection -> LLM Summarization
Final Prediction: User Prefs + Neighbor Prefs + Context -> LLM Recommendation

System Modules

Privacy Transmission Module

Perturb sensitive user data locally before uploading to the server

Model or implementation: Statistical perturbation algorithms (not a neural model)

Multitask Reflective Preference Extraction

Distill fine-grained user preferences from history into a Knowledge Base

Model or implementation: gpt-3.5-turbo

Neighbor Preference Retrieval

Inject collaborative signals by summarizing preferences of similar users

Model or implementation: gpt-3.5-turbo

Multitask Next POI Recommendation

Generate final POI recommendations using all extracted context

Model or implementation: gpt-3.5-turbo

Novel Architectural Elements

Reflective Knowledge Base construction: Iteratively updating a preference profile via self-correction on history segments before inference
Neighbor Preference Summarization: Explicitly prompting the LLM to synthesize 'collaborative' signals from retrieved neighbor profiles

Modeling

Base Model: gpt-3.5-turbo

Reproducibility

Code not provided. Datasets (Foursquare SIN/NY/PHO) are standard benchmarks. Hyperparameters for differential privacy and prompt structures are detailed in the paper.

📊 Experiments & Results

Evaluation Setup

Next POI recommendation on chronological train/val/test splits (8:1:1)

Benchmarks:

Foursquare Singapore (SIN) (Next POI Recommendation)
Foursquare New York (NY) (Next POI Recommendation)
Foursquare Phoenix (PHO) (Next POI Recommendation)

Metrics:

ACC@K (Accuracy at K)
MRR (Mean Reciprocal Rank)
Statistical methodology: Repeated each evaluation 10 times and reported the average result

Key Results

Benchmark	Metric	Baseline	This Paper	Δ
MRP-LLM (without privacy, marked as MR-LLM) significantly outperforms existing LLM-based baselines across all datasets.
PHO	ACC@1	0.5350	0.6250	+0.0900
SIN	MRR	0.5474	0.6138	+0.0664
Even with full privacy protection enabled (MRP-LLM), the method maintains competitive performance against non-private baselines.
SIN	ACC@1	0.6000	0.5600	-0.0400
PHO	MRR	0.5829	0.6033	+0.0204
SIN	ACC@1	0.50	0.56	+0.06

Experiment Figures

Ablation study results (ACC/MRR) on three datasets when removing specific components (Preference Probing, Self-Reflection, Neighbor Retrieval, Privacy)

Main Takeaways

MRP-LLM consistently outperforms other LLM-based methods (LLMMob, LLMMove) across all datasets, validating the reflection and neighbor retrieval components
The method is robust to privacy perturbations; the performance drop when enabling differential privacy is relatively small (avg 4.5% ACC drop), remaining competitive with SOTA
Geographical neighbors provide the strongest collaborative signal compared to semantic or social neighbors
Multitask preference probing (extracting category/region/distance separately) is more effective than direct prediction, with category probing being the most impactful

📚 Prerequisite Knowledge

Prerequisites

Point-of-Interest (POI) Recommendation
Large Language Models (In-Context Learning)
Differential Privacy (DP)

Key Terms

POI: Point-of-Interest—a specific location (e.g., a restaurant or park) that a user visits

Differential Privacy: A mathematical framework for privacy that adds noise to data so that the output cannot reveal whether any specific individual's data was included

OUE: Optimized Unary Encoding—a local differential privacy mechanism that perturbs binary vectors (like one-hot category encodings) to protect data while maintaining statistical utility

Chain-of-Thought: A prompting technique where the LLM is encouraged to generate intermediate reasoning steps before the final answer

LBSN: Location-Based Social Network—platforms like Foursquare or Yelp where users check in to locations

KL divergence: Kullback-Leibler divergence—a statistical distance measure used here to quantify the similarity between two users' check-in distributions

CoT: Chain-of-Thought—a prompting strategy that encourages the model to 'think' step-by-step

Self-reflection: A process where the model critiques or updates its own previous outputs based on new evidence (in this case, historical ground truth)