MAP CLASSIFICATION WITH A SIMILARITY MEASURE

摘要

In many methods of knowledge discovery, data mining and machine learning, similarities between objects are one of the most important factors. In this paper, we require a similarity measure to have (1) clear meaning of the quantity and (2) adaptability to target problems. We propose a similarity measure that yields maximum a posteriori classification. This similarity measure is based on a belief that the more possibility belonging to the same class, the more similar they are. Using this similarity measure, we show that posterior probability over the class for an example is derived by vote which is similar to k-nearest neighbor (abbrev. k-NN). We compare the proposal method and a direct method which induce posterior probability over classes for examples.