Pattern classification with class probability output network.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Science

Background:

• Classifier outputs often lack direct probabilistic interpretation.
• Accurate posterior probabilities are crucial for reliable soft classification decisions.

Purpose of the Study:

• To develop a post-processing method for calibrating classifier outputs.
• To enable classifier outputs to represent accurate posterior probabilities of class membership.

Main Methods:

• Proposed a novel post-processing technique called the class probability output network (CPON).
• Analyzed classifier output distribution using beta distribution parameters.
• Optimized beta and kernel parameters to improve the uniformity of beta cumulative distribution function (CDF) values.

Main Results:

• The CPON method effectively calibrates classifier outputs to provide accurate posterior probabilities.
• Simulations using Support Vector Machine (SVM) classifiers on UCI datasets showed significant performance improvements.
• CPON outperformed standard SVM, SVM-related classifiers, and other probabilistic scaling methods.

Conclusions:

• The proposed CPON method offers a statistically meaningful improvement in classification performance.
• CPON enables classifiers to provide reliable posterior probabilities for soft decision-making.
• This approach enhances the interpretability and reliability of machine learning classifier outputs.