We attempt to discover the role and relative value of labeled and unlabeled samples in reducing the probability of error of the classification of a sample based on the previous observation of labeled and unlabeled data. We assume that the underlying densities belong to a regular family that generates identifiable mixtures. The unlabeled observations, under the above conditions, carry information about the statistical model and therefore can be effectively used to construct a decision rule. When the training set contains an infinite number of unlabeled samples, the first labeled observation reduces the probability of error to within a factor of two of the Bayes risk. Moreover subsequent labeled samples yield exponential convergence of the probability of classification error to the Bayes risk. We argue that labeled samples are exponentially more valuable than unlabeled samples and identify the (exponent as the Bhatthacharyya distance.
[1]
H. Teicher.
Identifiability of Mixtures
,
1961
.
[2]
H. Teicher.
Identifiability of Finite Mixtures
,
1963
.
[3]
Earl E. Swartzlander,et al.
Introduction to Mathematical Techniques in Pattern Recognition
,
1973
.
[4]
Richard O. Duda,et al.
Pattern classification and scene analysis
,
1974,
A Wiley-Interscience publication.
[5]
Satish Chandra,et al.
On the Mixtures of Probability Distributions
,
1977
.
[6]
Thomas M. Cover,et al.
Elements of Information Theory
,
2005
.