RSSL (version 0.9.8)

MCPLDA: Maximum Contrastive Pessimistic Likelihood Estimation for Linear Discriminant Analysis

Description

Maximum Contrastive Pessimistic Likelihood (MCPL) estimation (Loog 2016) attempts to find a semi-supervised solution that has a higher likelihood compared to the supervised solution on the labeled and unlabeled data even for the worst possible labeling of the data. This is done by attempting to find a saddle point of the maximin problem, where the max is over the parameters of the semi-supervised solution and the min is over the labeling, while the objective is the difference in likelihood between the semi-supervised and the supervised solution measured on the labeled and unlabeled data. The implementation is a translation of the Matlab code of Loog (2016).

Usage

MCPLDA(X, y, X_u, x_center = FALSE, scale = FALSE, max_iter = 1000)

Arguments

X

matrix; Design matrix for labeled data

y

factor or integer vector; Label vector

X_u

matrix; Design matrix for unlabeled data

x_center

logical; Should the features be centered?

scale

logical; Should the features be normalized? (default: FALSE)

max_iter

integer; Maximum number of iterations

References

Loog, M., 2016. Contrastive Pessimistic Likelihood Estimation for Semi-Supervised Classification. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(3), pp.462-475.

See Also

Other RSSL classifiers: EMLeastSquaresClassifier, EMLinearDiscriminantClassifier, GRFClassifier, ICLeastSquaresClassifier, ICLinearDiscriminantClassifier, KernelLeastSquaresClassifier, LaplacianKernelLeastSquaresClassifier(), LaplacianSVM, LeastSquaresClassifier, LinearDiscriminantClassifier, LinearSVM, LinearTSVM(), LogisticLossClassifier, LogisticRegression, MCLinearDiscriminantClassifier, MCNearestMeanClassifier, MajorityClassClassifier, NearestMeanClassifier, QuadraticDiscriminantClassifier, S4VM, SVM, SelfLearning, TSVM, USMLeastSquaresClassifier, WellSVM, svmlin()