Bayes risk weighted VQ and learning VQ

This paper examines two vector quantization algorithms which can combine the tasks of compression and classification: Bayes risk weighted vector quantization (BRVQ) proposed by Oehler et al. (1991), and optimized learning vector quantization 1 (OLVQ1) proposed by Kohonen et al. (1988). BRVQ uses a parameter λ to control the tradeoff between compression and classification. BRVQ performance is studied for a range of λ values for four classification problems. Increasing the λ parameter in BRVQ is intended to improve classification performance. However, for two of the problems studied, increasing λ degraded classification performance. A majority rule reclassification of the final codebook (using only the training set) greatly improves high-λ BRVQ performance for these cases. Finally, we compare the classification performance and mean square error (MSE) performance of BRVQ to that of OLVQ1 for four classification problems. BRVQ with codebook reclassification is found to have a lower MSE than OLVQ1 while maintaining comparable, but slightly inferior, classification performance