Conformal Loss-Controlling Prediction.

Area of Science:

• Machine Learning
• Statistical Learning Theory

Background:

• Conformal prediction (CP) provides coverage guarantees for prediction sets.
• Existing methods focus on controlling average risk, not individual prediction losses.

Purpose of the Study:

• To extend conformal prediction (CP) to control general loss functions for individual test objects.
• To develop a framework for conformal loss-controlling prediction.

Main Methods:

• Proposed a novel learning framework: conformal loss-controlling prediction.
• Proved theoretical guarantees under data exchangeability for finite-sample cases.
• Empirically tested the framework on classification and regression problems.

Main Results:

• Demonstrated effective control over general loss functions for individual predictions.
• Validated the approach in class-varying loss classification and weather forecasting postprocessing.
• Theoretical analysis and experiments confirm the framework's effectiveness.

Conclusions:

• Conformal loss-controlling prediction offers a powerful extension to CP for controlling specific prediction losses.
• The framework is effective for diverse applications, including point-wise classification and regression.
• This work advances controllable prediction methods beyond average risk.