Abstract
Missing Point-of-Interest (POI) category identification is a critical task in Location-based Social Networks (LBSNs) for enriching incomplete check-in data. Unlike traditional next-POI prediction, which relies on unidirectional history, identifying a missing gap in a sequence requires analyzing both preceding and succeeding contexts. We present Bi-GTPPP, a novel neural network model that integrates bi-directional sequence features with global "common sense" transition patterns and individual user preferences. Extensive evaluations on real-world datasets from New York City and Tokyo demonstrate that Bi-GTPPP significantly outperforms state-of-the-art baselines in both identification accuracy and ranking quality.
Introduction
In Location-based Social Networks (LBSNs) like Foursquare or Yelp, check-in data provides a foundation for personalized recommendations. However, these sequences are often sparse or incomplete. While most existing research focuses on Next POI Prediction, the authors identify a distinct need for Missing POI Category Identification.
Consider a user at "Work" at 5:00 PM and a "Cinema" at 9:00 PM, with a missing 7:00 PM check-in. A unidirectional model only sees "Work," whereas a bi-directional approach recognizes the likely intermediate stop at a "Restaurant." The challenge lies in effectively integrating these bi-directional temporal cues with two distinct types of influence: global transition patterns (societal norms) and personal preferences (individual habits).
Methodology: The Bi-GTPPP Model
The proposed Bi-directional Global Transition Patterns and Personal Preferences (Bi-GTPPP) model utilizes a multi-layered architecture to capture complex mobility dynamics.
The model consists of several key components:
- Bi-directional LSTM: Two Long Short-Term Memory networks capture the "forward" context (pre-gap) and the "backward" context (post-gap).
- Global Transition Embeddings: The model learns two transition matrices representing the likelihood of moving between categories across the entire dataset, capturing general societal behaviors.
- Personal Preference Embeddings: A latent factor model learns unique preference profiles for every user based on historical check-in frequency.
- Attention Matching Cell: This core innovation calculates normalized cosine similarity between sequence features and preference features. It dynamically weights patterns based on how well the user's recent behavior matches known global or personal trends.
Experimental Setup
The model was evaluated on two large-scale Foursquare datasets: NYC (227,428 check-ins) and TKY (573,703 check-ins). Category information was masked in a portion of the check-ins to create a ground truth for testing identification performance.
Results and Discussion
Bi-GTPPP-nonstatic (initialized with user history) achieved superior performance across all metrics, as shown in the table below.
Key Findings
- Bi-directionality is Essential: Bi-directional models consistently outperformed unidirectional ones (e.g., standard LSTMs or GRUs), proving that destination context is as informative as origin context.
- Global vs. Local Synergy: While personal preferences are powerful, global patterns provide a necessary "prior," especially for users with sparse histories. For example, the model can fall back on the fact that most people in NYC go from "Office" to "Subway."
- Category vs. Location: Traditional spatio-temporal models (like Bi-STDDP) are less effective for category identification because categories are abstract and less dependent on physical distance than specific locations.
- Efficiency Gains: By identifying the top 10 most likely categories, Bi-GTPPP can reduce the search space for specific POI recommendations by approximately 25 times (from ~60,000 POIs to ~2,500).
Conclusion and Future Work
Bi-GTPPP demonstrates the efficacy of combining bi-directional modeling with global and personal contexts for POI category identification. Future research directions include expanding the application scope to urban planning and criminal activity analysis, improving model interpretability, and incorporating theories from nonlinear systems to further refine human mobility modeling.
Summary updated