Decision Confidence Assessment in Multi-Class Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Automated systems aim to reduce human workload in classification tasks.
• Current multi-class recognition often requires full human expert evaluation, which is time-consuming.
• Optimizing expert time is crucial for improving production processes.

Purpose of the Study:

• To develop a novel approach for assessing decision confidence in multi-class recognition.
• To provide a tool that reduces the workload of human experts by filtering samples.
• To enable systems to decide whether a sample requires manual evaluation based on classifier certainty.

Main Methods:

• Implementing a decision confidence assessment instead of hard classification.
• Evaluating classifier certainty for each processed example.
• Developing a system that flags samples for manual review when confidence is low.
• Adjusting the method for any number of classes and optimizing for accuracy or coverage.

Main Results:

• The proposed method successfully assesses decision confidence in multi-class recognition.
• Experimental results demonstrate the approach meets predefined quality criteria.
• The system effectively identifies samples that benefit from human expert intervention.

Conclusions:

• The novel approach offers a practical solution for managing expert workload in classification.
• Decision confidence assessment is a viable alternative to hard classification for optimizing expert time.
• The method is flexible and adaptable to various classification scenarios and user preferences.