UNIGE document Master
previous document  unige:102928  next document
add to browser collection
Title

Prediction Divergence Criterion for Model Selection in the Logistic Regression

Author
Miglioli, Cesare
Director
Denomination Master of Science in Statistics
Defense Maîtrise : Univ. Genève, 2018
Abstract In this Master Thesis, we have analytically derived and numerically implemented three estimators of the Prediction Divergence Criterion (Avella-Medina et al., working paper) for Model Selection within the logistic regression framework. After the validation of these estimators by means of simulations, we have performed Model Selection both when the order of the variables was known in advance and when the order was correct but decided by an already existing algorithm, namely the binary lasso (Friedman et al., 2010). Finally we have produced evidences of the good performance of two of these estimators, one derived from the L2 norm error measure and the other from the binomial deviance, respectively in highly and moderately correlated settings. They have been proven better, to the extension of the simulation study, than the defaults methods, based on 10-fold Cross Validation, currently available in the glmnet(Friedman et al., 2017) R package.
Keywords Model selectionLogistic regressionDivergenceBig dataHigh dimensionsBinary lassoCross-validationStopping criteria
Full text
Master (437 Kb) - public document Free access
Structures
Citation
(ISO format)
MIGLIOLI, Cesare. Prediction Divergence Criterion for Model Selection in the Logistic Regression. Université de Genève. Maîtrise, 2018. https://archive-ouverte.unige.ch/unige:102928

100 hits

18 downloads

Update

Deposited on : 2018-03-14

Export document
Format :
Citation style :