Local differentially private multi-criteria recommender system
Expert Systems with Applications, cilt.331, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 331
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.eswa.2026.133314
- Dergi Adı: Expert Systems with Applications
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Aerospace Database, Applied Science & Technology Source, Compendex, INSPEC, Public Affairs Index, Academic Search Ultimate (EBSCO), Engineering Source (EBSCO), Technology Collection (ProQuest)
- Anahtar Kelimeler: Collaborative filtering, Differential privacy, Laplace mechanism, Multi-criteria recommender system, Privacy-preserving recommendation
- Bilecik Şeyh Edebali Üniversitesi Adresli: Evet
Özet
While multi-criteria collaborative filtering (MCCF) enhances recommender systems with more accurate and personalized content, it also introduces significant privacy concerns due to the use of fine-grained personal data. Differential privacy (DP), particularly local DP (LDP), is widely used in single-criterion systems to protect privacy. However, LDP-based mechanisms remain unexplored in MCCF systems.This study introduces the first LDP-enabled MCCF framework that applies a criterion-wise perturbation procedure to preserve the structure of multi-dimensional ratings while providing formal privacy guarantees. Three Laplace-based mechanisms as Standard (SLM), Bounded (BLM), and Truncated (TLM) are integrated and evaluated on two Yahoo!Movies subsets under privacy budgets ε ∈ [0.1, 10]. The results show that BLM and TLM consistently outperform SLM, particularly under stricter privacy conditions. On YM20, the root mean squared error (RMSE) loss compared to the unmasked variant ranges from 3.92–140.62% with SLM, 3.92-69.58% with BLM, and 3.68–88.47% with TLM. For YM10, RMSE loss ranges from 1.72–130.12% with SLM, 2.09–63.25% with BLM, and 1.55–75.64% with TLM. Moreover, minimizing noise via min-max normalization generally enhances accuracy, while rating distributions and sparsity of the datasets, clipping strategies and privacy levels further affect outcomes. Overall, the proposed LDP-based MCCF framework achieves competitive prediction quality while ensuring strong privacy guarantees, advancing privacy-preserving approaches in multi-criteria recommendation.