Conformal region classification focuses on developing region classifiers; i.e., classifiers that output regions (sets) of classes for new test instances. 2,13,16 Conformal region classifiers have been proven to be valid for any significance level epsilon is an element of [0, 1] in the sense that the probability the class regions do not contain the true instances' classes does not exceed e. In practice, however, conformal region classifiers need to be also efficient; i.e., they have to output non-empty and relatively small class regions. In this paper we show that conformal region classification can benefit from instance-transfer learning. Our new approach consists of the basic conformal region classifier with a nonconformity function that implements instance transfer. We propose to learn such a function using a new multi-class Transfer AdaBoost. M1 algorithm. The function and its relation to the conformal region classification are theoretically justified. The experiments showed that our approach is valid for any significance level epsilon is an element of [0, 1] and its efficiency can be improved with instance transfer.
|Journal||International Journal on Artificial Intelligence Tools|
|Publication status||Published - 2015|