Date(s) - 05/13/2016 - 05/14/2016
Categories No Categories
Start Date: 5/13/2016
End Date: 5/14/2016
Temple University Center City
1515 Market Street
Taught byStephen Vardeman, Ph.D.
Modern researchers increasingly find themselves facing a new paradigm where data are no longer scarce and expensive, but rather abundant and cheap. Both numbers of cases/instances and numbers of variables/features are exploding. This new reality raises important issues in effective data analysis.
Of course, the basic statistical objectivediscovery and quantitative description of simple structureremains unchanged. But new possibilities for applying highly flexible methods (not practical in small data contexts) must be reconciled with the inherent sparsity of essentially any data set comprised of a large number of featuresand the corresponding danger of overfitting and unwarranted generalization from data in hand. Modern statistical machine methods rationally and effectively address these new realities.
This course first describes and explains the new context, formulates issues that it raises, and points to cross-validation as a fundamental tool for matching method flexibility/complexity to data set information content in predictive problems. Then a variety of modern squared error loss prediction methods (modern regression methods) will be discussed, related to optimal prediction, and illustrated using standard R packages. These will include
- smoothing methods
- shrinkage for linear prediction (ridge, lasso, and elastic net predictors)
- regression trees
- random forests, and
Next a variety of modern classification methods will be introduced, related to optimal classification, and illustrated using standard R packages. These will include:
- linear methods for classification (linear discriminant analysis, logistic regression, support vector classifiers)
- kernel extensions of support vector classifiers
- classification trees
- adaboost, and
- other ensemble classifiers
Finally, well discuss some methods of modern unsupervised statistical machine learning, where the object is not prediction of a particular response variable but rather discovery of relations among features or natural groupings of either cases or features. These will include principal components and clustering methods
The course will consist of both lectures and hands-on R sessions.