Source-linked AI summary
A Survey of Point-of-interest Recommendation in Location-based Social Networks
Shenglin Zhao, Irwin King, Michael R. Lyu
TL;DR
POI recommendation must account for spatial, temporal, social, and heterogeneous-information challenges that distinguish LBSNs from conventional recommendation settings. This survey systematically reviews the field through taxonomies of influential factors, methodologies, and tasks, along with datasets, metrics, and future directions. It concludes by organizing the literature into a panorama that highlights representative work and emerging directions.
Problem
POI recommendation faces spatial and temporal constraints, complex relations, and heterogeneous information beyond conventional recommendation settings.
Method
The survey reviews over 50 studies and classifies systems by influential factors, methodology, and recommendation task.
Results
The survey presents a structured account of geographical, social, temporal, and content influences, fused and joint methods, and general and successive recommendation tasks.
Takeaways & Limitations
The survey provides a balanced panorama of POI recommendation research and identifies ranking-based models and online recommendation as future directions.
Takeaways & Limitations
Social influence can provide limited improvement, and prior temporal models are unsatisfactory for the POI recommendation scenario.
Abstract
from arXiv · showhide
Point-of-interest (POI) recommendation that suggests new places for users to visit arises with the popularity of location-based social networks (LBSNs). Due to the importance of POI recommendation in LBSNs, it has attracted much academic and industrial interest. In this paper, we offer a systematic review of this field, summarizing the contributions of individual efforts and exploring their relations. We discuss the new properties and challenges in POI recommendation, compared with traditional recommendation problems, e.g., movie recommendation. Then, we present a comprehensive review in three aspects: influential factors for POI recommendation, methodologies employed for POI recommendation, and different tasks in POI recommendation. Specifically, we propose three taxonomies to classify POI recommendation systems. First, we categorize the systems by the influential factors check-in characteristics, including the geographical information, social relationship, temporal influence, and content indications. Second, we categorize the systems by the methodology, including systems modeled by fused methods and joint methods. Third, we categorize the systems as general POI recommendation and successive POI recommendation by subtle differences in the recommendation task whether to be bias to the recent check-in. For each category, we summarize the contributions and system features, and highlight the representative work. Moreover, we discuss the available data sets and the popular metrics. Finally, we point out the possible future directions in this area and conclude this survey.
1 Introduction
POI recommendation helps users discover locations by mining check-ins, venue information, and social relationships, but differs from traditional recommendation because physical constraints create spatial and temporal behavior, complex relations, and heterogeneous information. The survey organizes prior systems by influential factors, methodology, and task.
- Motivation: POI recommendation mines check-in records, venue information, and social relationships to recommend locations users may visit.It can help users discover places and provide merchants with opportunities to reach potential consumers.
- Challenges: Unlike conventional recommendation, POI recommendation is shaped by physical constraints that impose geographic and temporal limits on check-ins.Users visit geographically constrained areas, while shops operate during limited periods.
- Challenges: POI recommendation involves complex relations among locations, users, and geographically influenced social connections.Location sharing can alter user relationships, while locations introduce new relation types.
- Challenges: LBSNs combine check-ins, geographic information, venue descriptions, social relations, and media such as comments and tweets.This heterogeneous information describes user activity from multiple perspectives.
- Survey organization: The survey classifies POI recommendation systems by influential factors, methodology, and task.Its taxonomies cover geographical, social, temporal, and content factors; fused and joint models; and general versus successive recommendation.
2 Problem Definition
The paper formalizes POI recommendation around users’ check-in records and sequences. Given all users’ sequences, the task is to recommend each user a list of unvisited POIs.
- Definitions: A check-in records a user visiting a POI at a particular time.It is represented as the triple ⟨u, l, t⟩.
- Definitions: A check-in sequence collects a user’s POI visits together with their timestamps.For simplicity, the sequence can be represented only by its ordered POI locations.
- Recommendation task: POI recommendation mines check-in records to recommend users a list of unvisited POIs.The task is defined over the collected check-in sequence set for all users.
3 Taxonomy by Influential Factors
POI recommendation reflects four influential factors—geographical, temporal, social, and content-related—whose models address the spatial, temporal, relational, and preference information in check-in behavior. The survey reviews representative approaches for modeling these factors and their observed effects.
- Overview: Four factors shape POI recommendation: geographical influence, temporal dynamics, social relations, and content indications.The survey treats check-in activity as a synthesized decision influenced by heterogeneous information and physical constraints.
- Geographical Influence: Geographical models capture users’ tendency to visit POIs near previously visited locations through power-law, Gaussian, and kernel-density approaches.Representative models include power-law distance distributions, multi-center Gaussian models, and personalized distance distributions from kernel density estimation.
- Geographical Influence: Power-law methods learn distance-distribution parameters from observed check-in data and combine geographical influence with collaborative filtering for POI recommendation.The power-law formulation models POI co-occurrence probability as a function of distance, with parameters learned through transformed regression or Bayesian estimation.
- Geographical Influence: Gaussian methods model users’ activity around multiple centers, while genetic-based Gaussian mixtures address potentially unbalanced POI assignments among centers.The multi-center Gaussian model represents activity around locations such as home and office; a genetic-based variant is reported to outperform the MGM.
- Geographical Influence: Personalized kernel density estimation models each user’s geographical influence with an individual distance distribution rather than a common distribution.This approach is motivated by differences among users’ check-in-location distributions and uses a bandwidth-controlled Gaussian kernel.
- Social Influence: Social influence can restrict collaborative filtering to similar friends or enter probabilistic matrix factorization through social regularization, but its gains for POI recommendation are limited.The survey attributes this limitation to the physical separation between online friendships and the locations users actually visit.
- Temporal Influence: Temporal models developed for other recommendation settings are unsatisfactory for POI recommendation because POI activity has distinct temporal characteristics.The survey therefore reviews specialized systems proposed to incorporate temporal influence into POI recommendation.
- Content Indications: User comments provide explicit preference information that complements check-in behavior and can improve understanding of users’ location preferences.A check-in may indicate only a broad preference, whereas comments can express approval, dislike, or specific reasons for a location.
4 Taxonomy by Methodology
The survey classifies POI recommendation systems into fused and joint models. Fused models combine separate recommendation signals, whereas joint models learn user preferences and influential factors together, including geographical and temporal effects.
- Methodology taxonomy: POI recommendation systems are categorized as fused models or joint models based on how influential factors are incorporated.Fused models combine recommendations from collaborative filtering and factor-specific models; joint models learn preference and influential factors together.
- Fused models: Fused models typically combine collaborative filtering with geographical and temporal recommendations because social influence and comments provide limited or missing information.A representative fused framework multiplies preference estimation by geographical visiting probability.
- Fused models: PMF-based fused modeling learns user and POI latent features from check-in data, then combines preference scores with geographical influence.The PMF formulation uses check-in frequencies and indicators, while the geographical component estimates location-visiting probability.
- Joint models: Joint models connect multiple influential factors to the same check-in target, modeling check-in behavior as a synchronized decision.The survey presents joint models as more reflective of real scenarios than separately fused factors.
- MF-based joint models: LRT jointly models temporal effects by learning time-dependent user features and regularizing consecutive temporal states.It estimates check-in preference from user and POI latent representations, then aggregates contributions across temporal states with an aggregation function.
- MF-based joint models: GeoMF extends weighted regularized matrix factorization with user activity-area and POI influence-area representations.It treats check-ins as implicit feedback, learns latent features, estimates check-in possibilities, and recommends higher-scoring POIs.
- Generative graphical models: The survey also reviews geographical probabilistic factor analysis, which models user preferences, geographical influence, and mobility patterns in a generative graphical framework.Regions are represented with Gaussian geographical distributions for POI modeling.
5 Taxonomy by Task
The survey distinguishes general POI recommendation from successive POI recommendation. General recommendation ranks likely POIs for users, while successive recommendation additionally models recent-location transitions and temporal context.
- General POI recommendation: General POI recommendation ranks unvisited POIs and selects the top N candidates with the highest estimated check-in possibility.This task is analogous to top-N recommendation in conventional systems and can incorporate geographical and temporal influence.
- General POI recommendation: Geo-FM models check-in probability through user preference and geographical influence from neighboring POIs.Its score combines user-target-POI interaction with interactions involving the target POI’s k-nearest neighbors.
- General POI recommendation: Geo-FM learns POI rankings with WARP loss, treating higher check-in frequency as stronger user preference.The loss converts ranking incompatibility into an optimization objective over observed check-ins and candidate POIs.
- Successive POI recommendation: Successive POI recommendation uses the most recent checked-in location, requiring both user-preference modeling and POI-correlation analysis.It extends general recommendation by focusing on prompt recommendations conditioned on recent check-in context.
- Successive POI recommendation: STELLAR ranks candidate POIs with a time-aware score conditioned on the user, query POI, timestamp, and candidate POI.Its latent ranking framework models user preference, time effects, and correlation between the last and candidate POIs.
6 Performance Evaluation
The survey reviews benchmark datasets and metrics for evaluating POI recommendation. It emphasizes precision, recall, F-score, and relative metrics that compare performance with random recommendation.
- Datasets: Gowalla, Brightkite, and Foursquare are identified as benchmark LBSN datasets for evaluating POI recommendation models.The survey summarizes their statistics in a dataset table.
- Metrics: Precision and recall are the main general metrics used to evaluate POI recommendation results.Precision measures recovered POIs among recommendations, while recall measures recovered POIs among test-set POIs.
- Relative metrics: Relative precision@N and recall@N compare recommendation performance against random selection.The relative metrics use candidate-POI sets and random-recommendation baselines for comparison.
- Metrics: F-score combines precision and recall through their harmonic mean.It provides a single measure reflecting the balance between the two metrics.
7 Trends and New Directions
The survey identifies ranking-based modeling and online recommendation as directions for POI recommendation. Ranking methods target preference order, while online methods may address cold-start and changing user behavior.
- Ranking-based models: Ranking-based models learn users’ preference order rather than predicted check-in probability values.The survey notes that recommendation decisions depend on ordering, and several recent models use ranking objectives.
- Online recommendation: Online recommendation is proposed as a future direction because offline models struggle with new users and changing behavior.Online methods may improve adaptability, while the survey notes that online POI recommendation had not yet been used in the cited work.
8 Conclusion
The survey reviews over 50 recent POI recommendation papers and organizes them by influential factors, methodology, and task. It also identifies representative work, research trends, and possible new directions.
- The survey reviews over 50 POI recommendation papers from major conferences and journals.
- It categorizes POI recommendation systems by influential factors, methodology, and task.
- The survey reports representative work within each category.
- It presents a broad view of the field while identifying research trends and possible new directions.