Presupervised and post-supervised prototype classifier design

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• Nearest prototype classifiers are fundamental in pattern recognition.
• Existing methods often rely on hard assignments of prototypes to classes.

Purpose of the Study:

• To generalize the nearest prototype classifier using soft class labeling.
• To introduce and analyze presupervised and postsupervised GNPC designs.
• To determine conditions for optimality relative to the Bayes error rate.

Main Methods:

• Extension of the nearest prototype classifier to a generalized version (GNPC).
• Development of presupervised and postsupervised GNPC designs based on prototype selection.
• Derivation of optimality conditions for GNPC designs.
• Experimental validation using artificial and real-world datasets (satimage).
• Implementation using radial basis function (RBF) networks.

Main Results:

• Two GNPC designs were derived: presupervised (prototypes from class-conditional densities) and postsupervised (prototypes from unconditional density).
• Optimality conditions relative to Bayes error rate were established for both designs.
• Experimental results on artificial and satimage datasets did not show a clear preference for either approach.
• The postsupervised GNPC demonstrated higher robustness but lower accuracy compared to the presupervised GNPC.

Conclusions:

• The generalized nearest prototype classifier (GNPC) offers a versatile framework encompassing various classifiers through soft labeling.
• Both presupervised and postsupervised GNPC designs are theoretically sound, with no definitive advantage identified.
• The choice between presupervised and postsupervised GNPC may depend on the desired trade-off between robustness and accuracy.