Federated Learning for Data Trading Portfolio Allocation With Autonomous Economic Agents

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Data is a critical resource in ubiquitous intelligence societies.
• Autonomous Economic Agents (AEAs) are needed for intelligent data trading.
• Limited local information hinders AEA portfolio allocation model training.

Purpose of the Study:

• To develop a novel data trading market for AEAs with exclusive local information.
• To enable AEAs to jointly train robust portfolio allocation models using federated learning (FL).
• To enhance revenue return estimation by integrating Histogram of Oriented Gradients (HoGs) and Discrete Wavelet Transformation (DWT).

Main Methods:

• AEAs employ federated learning (FL) to collaboratively train portfolio allocation models.
• A novel combination of HoGs and DWT is used to represent non-stationary revenue data.
• Local model drifts in the transform domain are leveraged for efficient global model updates.

Main Results:

• The proposed schemes demonstrate superior performance in simulations.
• Key metrics such as test loss, cumulative return, portfolio risk, and Sharpe ratio are improved.
• Reduced communication burden and enhanced training efficiency were observed.

Conclusions:

• The integrated approach effectively addresses information constraints in data trading.
• The method provides a robust solution for optimizing AEA portfolio allocation.
• This research advances intelligent data trading in ubiquitous intelligence environments.