Learning with General Proximity Measures

Proximity is the basic quality which identifies and characterizes groups of objects in various domains and contexts. When objects are compared to a set of chosen prototype examples, proximity can be used as a natural ingredient to build a numerical representation. Pattern classes may be learned from such proximity representations by the traditional nearest neighbor rule, as well as by other alternative strategies. These encode the proximity information in suitable representation vector spaces in which statistical classifiers can be trained. Such recognition techniques can be successful, provided that the measure is informative, independently whether it is metric or Euclidean, or not.