Online Regression for Data With Changepoints Using Gaussian Processes and Reusable Models.

Area of Science:

• Machine Learning
• Statistical Modeling
• Data Science

Background:

• Gaussian Process (GP) models are crucial for prediction and control.
• Existing GP methods struggle with nonstationary data from multiple processes.
• Current nonstationary GP regression is computationally intensive and often not real-time.

Purpose of the Study:

• To develop an efficient online Gaussian Process framework for nonstationary data.
• To enable real-time changepoint detection and regression.
• To address computational and theoretical limitations of existing methods.

Main Methods:

• Introduced Gaussian Process Non-Bayesian Clustering (GP-NBC).
• Developed an efficient online framework for GP regression.
• Implemented real-time changepoint detection and regression capabilities.

Main Results:

• GP-NBC demonstrates superior performance on real-world and synthetic datasets.
• Achieved significant computational time reduction (e.g., 98% vs. Dirichlet process GP clustering).
• Maintained comparable mean absolute error to state-of-the-art methods.

Conclusions:

• GP-NBC offers an efficient and effective solution for nonstationary regression.
• The framework is suitable for real-time applications requiring dynamic data analysis.
• Provides provable guarantees on correctness and speed for online GP learning.