Distributed Nonparametric Regression with Heterogeneity Through Prediction-Based Aggregation

Area of Science:

• Distributed statistical modeling
• Machine learning
• Data privacy

Background:

• Large-scale datasets present challenges in statistical modeling, particularly regarding data privacy and communication efficiency.
• Heterogeneous distributed environments require adaptable modeling approaches.

Purpose of the Study:

• To propose a data-driven weighted aggregation procedure for distributed statistical modeling.
• To enhance communication efficiency and adaptability in heterogeneous environments.
• To ensure data privacy in large-scale data analysis.

Main Methods:

• A data-driven weighted aggregation procedure utilizing the squared prediction error matrix.
• Analysis of asymptotical optimal weights and corresponding risk.
• Investigation of the minimax property for nonparametric function estimates.
• Monte Carlo simulations for finite sample performance evaluation.

Main Results:

• The proposed estimates achieve asymptotical optimal weights concerning quadratic loss and risk.
• Limits of data-driven weights are derived.
• Demonstrated effectiveness through simulations and a real-world heart rate prediction dataset.

Conclusions:

• The developed weighted aggregation procedure is effective for distributed statistical modeling.
• The method ensures communication efficiency and adaptability in heterogeneous settings.
• It provides a robust approach for analyzing large-scale, privacy-sensitive data.