[Paper Review] Practical Privacy Preserving POI Recommendation

Chaochao Chen, Jun Zhou+5 more|arXiv (Cornell University)|Mar 5, 2020

Privacy-Preserving Technologies in Data66 references4 citations

TL;DR

This paper proposes PriRec, a privacy-preserving POI recommendation framework that keeps users' private data and linear models on their devices while using local differential privacy to securely collect perturbed user-POI interactions for generating dynamic POI popularity features. The framework combines secure decentralized gradient descent for user-side linear models and secure aggregation in federated learning for the feature interaction model, achieving comparable or better accuracy than centralized FM with strong privacy guarantees.

ABSTRACT

Point-of-Interest (POI) recommendation has been extensively studied and successfully applied in industry recently. However, most existing approaches build centralized models on the basis of collecting users' data. Both private data and models are held by the recommender, which causes serious privacy concerns. In this paper, we propose a novel Privacy preserving POI Recommendation (PriRec) framework. First, to protect data privacy, users' private data (features and actions) are kept on their own side, e.g., Cellphone or Pad. Meanwhile, the public data need to be accessed by all the users are kept by the recommender to reduce the storage costs of users' devices. Those public data include: (1) static data only related to the status of POI, such as POI categories, and (2) dynamic data depend on user-POI actions such as visited counts. The dynamic data could be sensitive, and we develop local differential privacy techniques to release such data to public with privacy guarantees. Second, PriRec follows the representations of Factorization Machine (FM) that consists of linear model and the feature interaction model. To protect the model privacy, the linear models are saved on users' side, and we propose a secure decentralized gradient descent protocol for users to learn it collaboratively. The feature interaction model is kept by the recommender since there is no privacy risk, and we adopt secure aggregation strategy in federated learning paradigm to learn it. To this end, PriRec keeps users' private raw data and models in users' own hands, and protects user privacy to a large extent. We apply PriRec in real-world datasets, and comprehensive experiments demonstrate that, compared with FM, PriRec achieves comparable or even better recommendation accuracy.

Motivation & Objective

• To address the privacy risks in centralized POI recommendation systems where both user data and models are held by the recommender.
• To design a scalable, efficient, and privacy-preserving recommendation framework that minimizes exposure of private user data and model parameters.
• To maintain high recommendation accuracy despite decentralizing sensitive data and models, ensuring practical deployment in real-world systems.
• To enable dynamic POI popularity features to be generated from perturbed user interactions without compromising user privacy.
• To develop a secure, efficient collaborative learning mechanism that preserves user-level personalization while protecting individual preferences.

Proposed method

• Users' private data (profiles and actions) are stored locally on their devices (e.g., smartphone), preventing central collection.
• Public POI data—static (e.g., category) and dynamic (e.g., visit count)—are maintained by the recommender to reduce client-side storage.
• Local differential privacy (LDP) is applied to perturb user-POI interaction data before transmission, enabling private aggregation of dynamic features.
• The recommendation model is based on Factorization Machine (FM), with linear components stored and trained locally via a secure decentralized gradient descent protocol.
• Feature interaction components are trained centrally using secure aggregation in a federated learning setting, preserving model privacy.
• A hybrid architecture separates data and model ownership: private data and linear models stay with users; shared feature interactions are trained securely on the server.

Experimental results

Research questions

• RQ1 Can a POI recommendation system preserve user data and model privacy without sacrificing recommendation accuracy?
• RQ2 How can dynamic POI popularity features be generated from perturbed user interactions while ensuring local differential privacy?
• RQ3 Can secure decentralized gradient descent enable effective collaborative learning of user-specific linear models without centralizing them?
• RQ4 Does the proposed PriRec framework achieve comparable or better performance than centralized FM models under strong privacy guarantees?
• RQ5 How efficient and scalable is the PriRec framework in practice, especially with limited online devices and communication?

Key findings

• PriRec achieves a 6.30% higher AUC than FM on the Koubei80 dataset when K=5, demonstrating superior or comparable accuracy despite decentralization.
• The integration of LDP-generated dynamic POI popularity features improves PriRec's AUC by 2.67% compared to PriRec- (without dynamic features) on Koubei90 with K=5.
• PriRec converges faster on Foursquare80 than Koubei80 due to fewer features (6 vs. 89), indicating faster learning on lower-dimensional data.
• Performance stabilizes with only 0.04% of users (30 out of 85,466) as maximum neighbors, proving scalability and practicality in sparse online environments.
• Training time scales linearly with data size, confirming the efficiency of PriRec and its feasibility for large-scale deployment.
• The framework maintains strong privacy guarantees by keeping raw user data and linear models on-device and using cryptographic techniques for secure aggregation.