Federated cross-view e-commerce recommendation based on feature rescaling.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Centralized recommendation systems raise data privacy issues.
• Federated learning trains models on client devices, preserving raw data privacy.
• Federated learning struggles with feature extraction efficiency and noise in e-commerce.

Purpose of the Study:

• Introduce Fed-FR-MVD, a multi-view federated learning framework.
• Enhance feature extraction efficiency and recommendation accuracy for e-commerce.
• Address data heterogeneity and noise sensitivity in federated recommendation systems.

Main Methods:

• Developed a novel multi-view federated learning framework (Fed-FR-MVD).
• Integrated a Feature Representation (FR) mechanism within a multi-view structure.
• Incorporated item and user perspectives for robust feature representation.
• Utilized dynamic rescaling to optimize feature utilization and mitigate noise.

Main Results:

• Achieved a 12%-18% increase in recommendation accuracy compared to existing methods.
• Demonstrated maintained performance across 5%-15% noise levels.
• Showcased improved resilience and efficiency in handling data heterogeneity.

Conclusions:

• Fed-FR-MVD offers a more resilient and efficient framework for federated recommendation systems.
• The framework effectively addresses privacy concerns in e-commerce environments.
• It enhances feature representation and robustness, crucial for data-diverse settings.