Suboptimal Bayesian classification by vector quantization with small clusters

摘要

Multi-dimensional classification based on the Bayes criterion minimizes the probability of misclassification. In order to apply this criterion, one has to know or to evaluate the probability densities of each class of data. Parzen windows or probabilistic neural networks may be used to estimate these probability densities; however, the number of operations involved in such process is prohibitive for large databases. The proposed algorithm shows how to apply vector quantization techniques to reduce the size of the learning set, while keeping sufficiently accurate estimations of probability densities. The problem of the width of the kernels used in the estimation is addressed by making the hypothesis of small clusters after quantization.