Privacy-Preserving Point-of-Interest Recommendation via Decentralized Matrix Factorization

The paper "Privacy Preserving Point-of-interest Recommendation Using Decentralized Matrix Factorization" was presented at AAAI 2018 by researchers from Ant Financial and Singapore University of Technology and Design. It proposes a decentralized matrix‑factorization method that keeps user interaction data on personal devices, thereby protecting privacy while reducing storage and computation waste.

Traditional POI recommendation systems are centralized: user‑item interaction data are collected on a server, leading to (1) excessive storage and compute resource consumption and (2) exposure of private user behavior to potential attacks.

To address these issues, the authors design a privacy‑preserving decentralized matrix‑factorization framework. Each user stores on their device: (a) raw POI interaction data, (b) their own user latent vector, (c) a shared (global) POI latent vector, and (d) a personalized (local) POI vector. Model training is performed locally, and users collaborate by exchanging information without revealing raw data.

The paper identifies two challenges: C1 – selecting which users should exchange information, and C2 – determining what information can be shared safely. For C1, the authors analyze geographic clustering of POI check‑ins, construct a user adjacency graph based on location proximity, and employ a Random Walk to select interaction partners. For C2, they propose exchanging gradients of the shared POI latent vectors, enabling collaborative model updates while preserving privacy.

Experiments were conducted on two datasets: the public Foursquare dataset and an internal Alipay dataset. Evaluation metrics include top‑k precision (P@k) and recall (R@k). Baselines include centralized matrix factorization (MF), Bayesian Personalized Ranking (BPR), and two ablated versions of the proposed model that retain only the global (GDMF) or only the local (LDMF) POI vectors. Results show that the decentralized matrix factorization consistently outperforms the baselines and that removing collaborative training (LDMF) leads to a significant performance drop.

Training and testing loss curves demonstrate stable convergence of the proposed method across iterations.

In conclusion, the decentralized approach offers a viable solution for privacy‑preserving POI recommendation, reducing resource waste and mitigating privacy risks. Future work will explore model compression techniques to address storage and computation constraints on user devices.

References: [1] Mnih, A., & Salakhutdinov, R. R. (2008). Probabilistic matrix factorization. In Advances in Neural Information Processing Systems (pp. 1257‑1264). [2] Rendle, S., Freudenthaler, C., Gantner, Z., & Schmidt‑Thieme, L. (2009). BPR: Bayesian personalized ranking from implicit feedback. In Proceedings of the 25th Conference on Uncertainty in Artificial Intelligence (pp. 452‑461). [3] Yan, F., Sundaram, S., Vishwanathan, S. V. N., & Qi, Y. (2013). Distributed autonomous online learning: Regrets and intrinsic privacy‑preserving properties. IEEE Transactions on Knowledge and Data Engineering, 25(11), 2483‑2493.