A Federated Recommendation System with a Dual-Layer Multi-Head Attention Network and Regularization Strategy.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Privacy

Background:

• Federated recommendation (FedRec) systems aim to balance personalized recommendations with user privacy.
• Existing FedRec models face challenges due to limited local user interaction data, hindering preference learning and increasing overfitting risk.

Purpose of the Study:

• To propose a novel federated recommendation system, FedDMR, that addresses the limitations of current FedRec models.
• To enhance the exploitation of interaction information and mitigate overfitting in decentralized recommendation scenarios.

Main Methods:

• FedDMR employs a dual-layer multi-head attention network to capture and enrich user and item embeddings from local interactions.
• A regularization strategy guides model updates, constraining deviation from global parameters to prevent overfitting and improve generalizability.
• The system utilizes a federated learning framework for decentralized training and parameter aggregation.

Main Results:

• FedDMR demonstrated an average improvement of 2.63% in AUC and precision across three datasets.
• The proposed model significantly outperformed recent federated recommendation baselines.

Conclusions:

• FedDMR effectively enhances personalized preference modeling in federated settings by enriching user feature representations.
• The dual-layer attention network and regularization strategy successfully mitigate overfitting and improve model generalizability, leading to superior recommendation performance.