Abstract
Point-of-interest (POI) recommendation suggests new places for users to visit, a task that has gained prominence with the rise of location-based social networks (LBSNs). This paper provides a systematic review of the field, exploring new challenges such as physical constraints that distinguish POI recommendation from traditional digital-only systems. We propose three taxonomies based on influential factors, modeling methodologies, and recommendation tasks. By summarizing contributions and highlighting representative work, we identify key trends, available datasets, and future research directions.
Introduction
Traditional recommendation systems, like those used by Netflix or Amazon, operate in environments where physical distance is irrelevant. In contrast, recommending a Point-of-Interest (POI)—such as a restaurant, park, or gym—introduces significant real-world complexities. Users are constrained by physical distance and time; a cafe 500 miles away is rarely a viable suggestion, and a bar is inaccessible at 10:00 AM if it is closed. This survey explores how researchers integrate these multi-dimensional realities—geography, time, and social ties—into mathematical models to bridge the gap between collaborative filtering and human mobility patterns.
Taxonomy of POI Recommendation
The survey organizes the existing literature into three distinct taxonomies to provide a structured understanding of the field.
1. Influential Factors
Modeling the "why" behind a check-in involves several key dimensions:
- Geographical Influence: This is the most significant factor. Researchers use Power Law distributions to model the preference for nearby places, Multi-center Gaussian Models for clustering around home or work, and Kernel Density Estimation (KDE) for personalized distance profiles.
- Social Influence: While friends often share interests, social influence is less effective in POI recommendation than in media recommendation due to geographical dispersion.
- Temporal Influence: Models must account for periodicity (daily routines), consecutiveness (sequential visits), and non-uniformness (weekend vs. weekday behavior).
- Content Indications: Natural Language Processing (NLP) is applied to user tips and sentiment analysis to refine preference scores.
2. Methodologies
How these factors are combined defines the modeling approach:
- Fused Models: These build separate models for each factor and combine their outputs (e.g., multiplying a preference score by a geographical probability).
- Joint Models: These use a single objective function to learn all factors simultaneously, often employing Matrix Factorization (MF) variants or Generative Graphical Models.
3. Recommendation Tasks
The goal of the system can vary:
- General POI Recommendation: Providing a static "Top-N" list of new places for a user.
- Successive POI Recommendation: Predicting the next location based on the user's current context (e.g., the STELLAR model).
Key Insights: The Role of Geography
Geographical modeling remains the primary driver of accuracy in POI systems. Empirical data shows a strong correlation between check-in probability and physical distance, typically following a Power Law distribution where the likelihood of a visit decreases as distance increases.
A notable finding is the "Social Distance Paradox." In digital domains, social ties are paramount. However, in LBSNs, physical proximity often trumps social similarity; a user is more likely to visit a nearby restaurant they haven't heard of than a highly-rated one in a different city recommended by a friend.
Datasets and Performance Trends
The research community relies on several "gold standard" datasets from platforms like Foursquare, Gowalla, and Brightkite. These datasets are unique because they pair user IDs with precise timestamps and GPS coordinates.
Analysis of these datasets reveals that Joint models generally outperform fused models by capturing latent interactions between factors. Furthermore, Successive recommendation models show striking improvements in Recall metrics by utilizing the user's current location as a "query POI."
Future Directions
The survey identifies several frontiers for future research:
- Cold-Start Problem: Developing effective strategies for new users with minimal check-in history.
- Online Recommendation: Moving from "offline" training on historical data to real-time adaptation using algorithms like multi-armed bandits.
- Ranking-based Refinement: Integrating "Learning to Rank" techniques (e.g., BPR or WARP loss) to prioritize the order of recommendations over absolute probability values.
Summary updated