Research on knowledge graph-based personalized recommendation algorithms and privacy protection in distributed networks

To improve the personalized recommendation performance of online platforms in a distributed environment while also considering user privacy protection, this paper proposes a hyperbolic attention knowledge graph recommendation algorithm for distributed architectures, integrating contextual information and privacy protection mechanisms.  First, In the design of the recommendation model, a hyperbolic attention model is employed to organically integrate semantic features with an attention mechanism, thereby fully capturing the inter-relationships between products. Next, a context-aware model is designed by combining an LSTM-based memory model with AFM-based feature crossing. Furthermore, to address the aforementioned security threats, differential privacy techniques are introduced to inject noise into user features, and collaborative model training across multiple nodes is implemented within a federated learning framework. Experimental results show that, under distributed and privacy-preserving constraints, knowledge graph recommendation algorithms can achieve better precision, recall, ACC, and AUC under the same conditions. Comparing our algorithm with other algorithms, our proposed algorithm improves the ACC by 2.9 to 11 percentage points and the AUC by 3.2 to 9.1 percentage points. Similarly, our proposed algorithm achieves RMSE and MAE of 0.960 and 0.858, respectively, which can demonstrate a significant advantage. It also performs well in terms of privacy protection (such as ε-differential privacy parameters) and communication efficiency. Furthermore, system security evaluation experiments demonstrate that when the privacy budget ε ≥ 2.0, the model effectively reduces the success rate of membership inference attacks and significantly degrades the reconstruction quality of gradient inversion attacks; it also exhibits a mitigating effect against model poisoning attacks under certain conditions, thereby validating the algorithm's security and practicality in real-world distributed recommendation scenarios. Comparative results across the various algorithms show that the proposed algorithm remains the best among all models in terms of precision and recall.