Meta-learning Enhanced Next POI Recommendation by Leveraging Check-ins from Auxiliary Cities

Jinze Wang, Lu Zhang, Zhu Sun, Yew-Soon Ong

Abstract

Next Point-of-Interest (POI) recommendation is a fundamental task in location-based social networks (LBSNs), yet it is frequently hindered by extreme data sparsity in target cities. This paper presents MERec (Meta-learning Enhanced next POI Recommendation), a framework that addresses this challenge by transferring behavioral knowledge from data-rich auxiliary cities. MERec utilizes a dual-channel encoder to capture both universal category-level transitions and city-specific POI sequences. A key innovation is the city-level correlation strategy, which weights knowledge transfer based on the behavioral similarity between cities. Experimental evaluations on Foursquare data from four diverse cities demonstrate that MERec consistently outperforms state-of-the-art baselines, providing significant gains in recommendation accuracy and ranking quality.

Introduction

Predicting a user's next destination based on their historical check-in sequence is essential for personalized services in location-based social networks. However, the effectiveness of these systems is often limited by the "data sparsity" problem. In many urban environments, the density of user-POI interactions is remarkably low; for instance, in New York City, interaction density can drop to 0.05%. While existing research has integrated spatial and temporal contexts to mitigate this, the potential of cross-city knowledge transfer remains largely untapped.

The primary obstacle to cross-city transfer is the lack of overlapping POIs—a specific venue in one city does not exist in another. Furthermore, "blind" transfer can lead to negative transfer, where irrelevant patterns from a culturally or geographically distinct auxiliary city degrade the model's performance in the target city. To overcome these hurdles, we propose MERec, a meta-learning framework designed to identify and leverage relevant behavioral patterns across urban boundaries.

The MERec Framework

Dual-Channel Encoding

MERec processes check-in data through two distinct channels to balance universal patterns with local specificities:

  1. Category-Level Channel: This channel focuses on transitions between POI categories (e.g., moving from a "Shopping Mall" to a "Restaurant"). Because category taxonomies are consistent across cities, they serve as a bridge for knowledge transfer. This channel is trained using a meta-learning paradigm (Model-Agnostic Meta-Learning, or MAML) to learn a generalized initialization of transition behaviors.
  2. POI-Level Channel: To capture the unique geography of the target city, this channel employs a Long Short-Term Memory (LSTM) network to model specific POI-to-POI transitions that are unique to the local environment.

City-Level Correlation Strategy

Not all auxiliary cities are equally useful for training a target model. MERec introduces a correlation strategy that calculates the Pearson correlation between cities based on their category-level behavioral distributions. During the global update phase of the meta-learning process, the model assigns higher weights to gradients originating from auxiliary cities that exhibit high correlation with the target city. This ensures that the model prioritizes knowledge from cities with similar cultural or transit behaviors (e.g., favoring Phoenix data when predicting for New York City over Singapore data).

Freezing and Fine-Tuning

To refine the model for the target city, MERec employs a two-step adaptation process. First, the early layers of the category encoder are frozen to preserve the generalized knowledge acquired from auxiliary cities. Subsequently, the remaining layers are fine-tuned using the limited data available in the target city, preventing the model from being overwhelmed by auxiliary noise while maximizing local relevance.

Experimental Evaluation

Datasets

The framework was evaluated using real-world check-in data from Foursquare across four cities with varying levels of data density:

  • Calgary (CAL): 1.06% density.
  • Phoenix (PHO): 0.22% density.
  • Singapore (SIN): 0.12% density.
  • New York (NYC): 0.05% density.

Performance Metrics

MERec was compared against seven baselines, including traditional collaborative filtering (BPRMF), deep learning models (ATST-LSTM), and existing meta-learning approaches (CHAML). Performance was measured using Hit Rate (HR@K) and Normalized Discounted Cumulative Gain (NDCG@K).

Results and Discussion

MERec achieved superior performance across all tested cities. On average, the framework demonstrated a 6.3% improvement in HR@K and a 6.53% improvement in NDCG@K compared to the strongest baselines.

Key findings from the study include:

  • The Importance of Correlation: Ablation studies confirmed that the city-level correlation strategy is vital. Without it, the model's performance dropped, indicating that indiscriminate knowledge transfer introduces significant noise.
  • Cross-City Similarity: Data analysis revealed that behavioral patterns in NYC are more closely aligned with Phoenix than Singapore. MERec successfully captured these nuances, validating the intuition that geographical and cultural proximity influences transit patterns.
  • Robustness to Sparsity: The most significant gains were observed in the most sparse datasets (NYC and SIN), proving MERec’s efficacy in scenarios where traditional models fail due to lack of data.

Conclusion

By framing next POI recommendation as a meta-learning problem and introducing a selective transfer mechanism, MERec effectively overcomes the data sparsity challenge. The use of category-level transitions as a universal language, combined with a correlation-based weighting system, allows for the successful migration of knowledge across cities. Future work will explore the integration of additional features, such as real-time weather and social network ties, to further enhance recommendation precision.

Summary updated

BibTeX
@inproceedings{wang2023meta,
  author = {Jinze Wang and Lu Zhang and Zhu Sun and Yew-Soon Ong},
  title = {{Meta-learning Enhanced Next POI Recommendation by Leveraging Check-ins from Auxiliary Cities}},
  booktitle = {PAKDD},
  year = {2023},
  doi = {10.1007/978-3-031-33380-4\_25},
}
Figure