Ensemble Predictors: Possibilistic Combination of Conformal Predictors for Multivariate Time Series Classification

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Conformal predictors (CP) offer a rigorous framework for uncertainty quantification in machine learning.
• Ensemble methods are widely used to improve model performance and robustness.
• Existing research has not fully explored the theoretical properties of CP ensembles.

Purpose of the Study:

• To propose a conceptual framework, Ensemble Predictors (EP), for studying ensembles of conformal predictors (CP).
• To investigate the theoretical properties of CP ensembles using possibilistic combination rules.
• To demonstrate the practical applicability and performance of EP in multivariate time-series classification.

Main Methods:

• Developed a conceptual framework for Ensemble Predictors (EP).
• Applied imprecise probabilities and possibilistic combination rules to CP ensembles.
• Evaluated EP on multivariate time-series classification benchmarks from the UCR archive.

Main Results:

• EP exhibits improved robustness, conservativeness, and accuracy in time-series classification.
• EP demonstrates competitive running times compared to standard algorithms.
• The proposed framework provides a novel approach to CP ensemble analysis.

Conclusions:

• Ensemble Predictors (EP) offer a theoretically grounded and practically effective approach for CP ensembles.
• The framework advances the understanding of CP ensembles in machine learning.
• EP methods show significant advantages for multivariate time-series classification tasks.