Adaptive relevance matrices in learning vector quantization.

Area of Science:

• Machine Learning
• Pattern Recognition
• Data Mining

Background:

• Relevance Learning Vector Quantization (RLVQ) is an efficient prototype-based classification algorithm.
• Existing RLVQ variations often use a weighted Euclidean metric, which may not fully capture complex feature relationships.

Purpose of the Study:

• To propose a new matrix learning scheme to extend RLVQ towards a general adaptive metric.
• To incorporate feature correlations and importance into the classification distance measure.
• To enable general metric adaptation during the training process.

Main Methods:

• Introduction of a full matrix of relevance factors into the distance measure.
• Automated consideration of correlations between features and their importance.
• Application of large margin generalization bounds independent of input dimensionality.
• Development of local metrics for piecewise quadratic decision boundaries.

Main Results:

• The proposed full matrix metric is more powerful than weighted Euclidean metrics for representing data structure.
• Generalization bounds independent of input dimensionality were achieved.
• The algorithm demonstrated effectiveness on artificial, benchmark, and bioinformatics datasets.

Conclusions:

• The new matrix learning scheme significantly enhances RLVQ by providing a more general and adaptive metric.
• This approach offers improved data representation and classification performance.
• The method is applicable to diverse problems, including bioinformatics tasks like splice site recognition.