Year: 2026
Type: article
Abstract: Recommender systems are widely used to support personalized item discovery; however, conventional centralized architectures still raise important concerns related to privacy exposure, reliance on a central authority, limited verifiability, data manipulation risks, and single-point-of-failure limitations.To address these issues, this research proposes a privacy-aware, blockchain-supported decentralized recommendation framework with off-chain recommendation computation and blockchain-based authorization and integrity verification, connecting Decentralized Communities-Deep Matrix Factorization (DC-DMF) with trusted authorized node selection, smart-contract-based announcement and finalization, and hash-based recommendation-delivery verification.The framework uses previously detected interest communities as the structural basis for community-aware recommendation computation.For each interest community, a trusted authorized node Ak is selected off-chain using social-relationship evidence from the social graph and then announced, confirmed, and finalized through a smart contract.Model training, recommendation ranking, and user-specific recommendation-package preparation are performed off-chain, while blockchain is used for authorization checks, compact hash-record storage, and delivery-integrity verification.The proposed framework was evaluated using the Douban Movie and Last.fm datasets under DMF-Whole network, CC-DMF, and DC-DMF recommendation scenarios.The results show that the voting stage was evaluated under different levels of available social-relationship evidence.Douban Movie provided limited social-relationship evidence, whereas Last.fm provided wider social-relationship evidence; however, both datasets produced finalized Ak values through the blockchain-based announcement and confirmation process.In Last.fm,DC-DMF mean Recall increased from 0.2915 at K=50 to 0.5588 at K=300, while CC-DMF mean Recall increased from 0.3473 to 0.6254.In Douban Movie, DC-DMF reported mean F1 values of 0.3129 and 0.3563 at K=10 and K=20, respectively, while CC-DMF reported mean F1 values of 0.2168 and 0.2267.In the recommendation-delivery stage, the stored delivery hashes matched the recomputed hashes of the corresponding off-chain recommendation packages, and the modification check showed that changing one recommended item in a representative package from each dataset resulted in a hash mismatch and rejection during verification.Overall, the results show that DC-DMF recommendation computation, blockchain-based community authorization, and hash-based delivery-integrity checking were integrated within the proposed on-chain/off-chain framework.
Institutions: University of Baghdad, University of Babylon
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7235341947
