Formulate Informed Initialization and Develop Customized Metrics for Prototypes of VQ

This paper presents an algorithm of vector quantization (VQ) that is encoded with the informed initialization of prototypes and the employment of prototype-customized metrics, and the algorithm is named as IMVQ. IMVQ extracts points of significant meaning from dataset and derives the first population of prototypes based on them. It creates 1-vs-rest SVMs model and uses its decision functions to learn individual metric for each prototype. The prototype updating takes two fashions rather than a single manner that is used in classical VQ. Experiments are conducted on real datasets to check the quality of initial prototypes, customized metrics as well as IMVQ and show the fine performance of the proposed strategies and algorithm.