How to solve the problem of distinguishing the
minority spectra from the majority of the spectra is quite im-
portant in astronomy. In view of this, an unbalanced spectra
classification method based on entropy (USCM) is proposed
in this paper to deal with the unbalanced spectra classifi-
cation problem. USCM greatly improves the performances
of the traditional classifiers on distinguishing the minority
spectra as it takes the data distribution into consideration in
the process of classification. However, its time complexity is
exponential with the training size, and therefore, it can only
deal with the problem of small- and medium-scale classifi-
cation. How to solve the large-scale classification problem is
quite important to USCM. It can be easily obtained by math-
ematical computation that the dual form of USCM is equiv-
alent to the minimum enclosing ball (MEB), and core vec-
tor machine (CVM) is introduced, USCM based on CVM is
proposed to deal with the large-scale classification problem.
Several comparative experiments on the 4 subclasses of K-
type spectra, 3 subclasses of F-type spectra and 3 subclasses
of G-type spectra from Sloan Digital Sky Survey (SDSS)
verify USCM and USCM based on CVM perform better
than kNN (k nearest neighbor) and SVM (support vector
machine) in dealing with the problem of rare spectra mining
respectively on the small- and medium-scale datasets and
the large-scale datasets.

EB00000003883K

Available