A fast conformal predictive system with regularized extreme learning machine.

Area of Science:

• Machine Learning
• Statistical Learning Theory
• Conformal Prediction

Background:

• Conformal predictive systems (CPS) provide statistically valid cumulative distribution functions (CDFs) and prediction intervals for regression.
• Existing CPS methods face computational challenges, limiting their application in time-sensitive domains like financial and weather forecasting.
• The validity property of CPS ensures statistical compatibility with observed data, crucial for risk-sensitive applications.

Purpose of the Study:

• To develop a computationally efficient conformal predictive system (CPS).
• To enhance the speed of CPS while maintaining statistical validity for regression tasks.
• To introduce a novel CPS algorithm that overcomes the inherent computational limitations of traditional methods.

Main Methods:

• Proposed a new CPS by combining leave-one-out cross-conformal predictive system (Leave-One-Out CCPS) with regularized extreme learning machine (RELM), termed LOO-CCPS-RELM.
• Analyzed the computational complexity of the proposed LOO-CCPS-RELM algorithm.
• Proved the asymptotic validity and controlled error rate of prediction intervals generated by LOO-CCPS-RELM under regularity assumptions.

Main Results:

• The LOO-CCPS-RELM method demonstrates improved computational efficiency compared to existing CPS approaches.
• Empirical validation across 20 public datasets confirmed the statistical validity of LOO-CCPS-RELM.
• The proposed method achieved favorable performance compared to other conformal predictive systems in experimental tests.

Conclusions:

• The LOO-CCPS-RELM offers a computationally efficient and statistically valid solution for regression problems.
• This advancement makes CPS more practical for real-world applications requiring fast and reliable predictions.
• The study validates the effectiveness of integrating RELM with CCPS for enhanced predictive system performance.