Abstract
Tree-structured vector quantizers for lossy data compression can be designed by combining clustering techniques with tree-structured methods for classification and regression as are developed in the statistics literature. Compression, on the one hand, and classification or regression, on the other, have differed primarily in the measures of quality and complexity used in the optimization algorithms. Given the similarity of the methods it is natural to consider combinations incorporating both squared error and Bayes risk into the design algorithms in order to simultaneously compress and classify local features accurately. We consider recent results of this type and compare them with other methods including independent design of classifier and compressor and Kohonen's (1989) "likelihood vector quantization"(LVQ).<>