Guaranteed Coverage Prediction Intervals With Gaussian Process Regression.

Area of Science:

• Machine Learning
• Statistical Modeling

Background:

• Gaussian Process Regression (GPR) offers uncertainty estimates but relies on model specification.
• Real-world applications often violate GPR's model specification assumption.
• This leads to unreliable prediction intervals (PIs) with inaccurate coverage levels.

Purpose of the Study:

• To develop an extension of GPR that ensures valid prediction interval coverage.
• To address the issue of misleading uncertainty estimates in GPR.
• To combine GPR's predictive power with Conformal Prediction's (CP) coverage guarantees.

Main Methods:

• An extension of GPR was developed using the Conformal Prediction (CP) framework.
• The proposed method integrates GPR with CP to ensure valid coverage.
• Experimental results were used to evaluate the performance against existing methods.

Main Results:

• The Conformal Prediction extension guarantees valid prediction intervals.
• This guarantee holds even when the GPR model is misspecified.
• Experimental results demonstrated the superiority of the proposed approach over existing methods.

Conclusions:

• The proposed GPR extension with CP provides reliable uncertainty estimates.
• This approach overcomes the limitations of standard GPR in practical applications.
• It offers a robust solution for accurate prediction intervals in machine learning.