Probabilistic Safety Regions via Finite Families of Adjustable Classifiers

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Supervised classification models identify patterns but inherently contain misclassification errors due to numerical approximations.
• Current error control in machine learning (ML) is often heuristic, with trade-offs between minimizing errors for one class at the expense of others.

Purpose of the Study:

• To develop theoretical foundations for probabilistic certifications of machine learning classifiers.
• Introduce a method for probabilistically controlling misclassification errors within specific input space regions.

Main Methods:

• Introduce the concept of a probabilistic safety region to define areas with controlled misclassification rates.
• Utilize adjustable classifiers, controllable by a scalar parameter, to link ML tuning with error control.

Main Results:

• Demonstrate a novel approach for providing probabilistic guarantees on classifier performance.
• Successfully applied the method to synthetic data, benchmark datasets, and a smart mobility application.

Conclusions:

• The proposed framework offers a theoretically grounded method for probabilistic error control in supervised classification.
• This approach enhances the reliability and trustworthiness of machine learning models by providing quantifiable safety guarantees.