Deck 6: Modeling and Machine Learning

Question 1

In one sentence, what does R's "caret" package do? Name three of its procedures discussed in Chapter 6.

Answer 1

The "caret" package in R supports comparison of 238 machine modeling approaches, including support vector machine (SVM) regression, gradient boosting machine (GBM), and learning vector quantization (LVQ) models.

Question 2

The support vector machine (SVM) method may be used to address which of the following?

A) Assembly line process problems
B) Classification problems
C) Regression problems
D) Both B and C

Answer 2

D

Question 3

In what package is the svm command found as the default SVM method?_______________________

Answer 3

It is found in the "e1071" package.
The "caret" package by default uses the ksvm command, which is from the "kernlab" package.

Question 4

Which is NOT a positive aspect of SVM methods?

A) It handles both classification and regression problems
B) It supports traditional parametric significance testing
C) Cross-validation is built in.
D) It handles nonlinearity and interaction effects autormatically

Question 5

Which is NOT a negative aspect of SVM methods?

A) SVM has "black box" aspects which make it less transparent than OLS regression
B) Trial and error methods may be needed to optimize the model
C) In spite of cross-validation it is still possible to overfit SVM models to noise in the data
D) All of the above.

Question 6

What does the SVM algorithm attempt to optimize?

A) Percent of variance explained
B) Number of Hyperplanes
C) Maximum margins
D) Maximum number of support vectors
E) Minimum cost

Question 7

A "loss function" is a metric to measure model performance. Is the loss function in SVM closer to that in OLS regression or logistic regression? Is it the same as the one you pick (explain briefly)?
Which closer? ______________________
Same?/Explain:

Question 8

What is the purpose of "kernels" in SVM?

A) Kernels try to linearize nonlinear problems and make a solution possible.
B) Kernels optimize model selection (selection of the best variables)
C) Kernels visualize SVM results
D) Kernels are the source code underlying SVM

Question 9

What is the default kernel in SVM regression using the svm command, and in most other SVM commands?

A) Linear
B) Polynomial
C) Sigmoid
D) Radial

Question 10

In SVM, what are gamma, degree, coef0, and nu?

A) Possible tuning parameters which may optimize the SVM model.
B) Labels for coefficients output by SVM as part of the solution
C) Labels for stages in solving a problem with SVM
D) All of the above

Question 11

SVM routinely outperforms OLS regression when the data a nonlinear?

Question 12

What is true of SVM in relation to understanding the relative importance of predictor variables?

A) SVM is a "black box" methodology which hides the relative roles of predictor variables.
B) Listing of variables by importance is a menu choice in RStudio.
C) Variable importance may be assessed through a leave-one-out method in which one predictor at a time is dropped from the model.
D) As with OLS regression, beta weights indicate relative predictive importance in SVM

Question 13

The author of the svm command believes SVM models must be tuned to obtain sensible results.

Question 14

When machine learning models (not just SVM) are implemented using the train command of the caret package, tuning of the model is accomplished on a built-in basis.

Question 15

For problems where the DV is binary, both OLS regression and SVM are inappropriate and logistic regression should be used instead.

Question 16

Which is NOT true of gradient boosting machine (GBM) models?

A) GBM can be implemented under the "caret" package.
B) Decision trees and random forests can be base learning methods under GBM.
C) OLS regression can be a base learner model under GBM.
D) The GBM solution is attained in an iterative process weighting cases differentially, with gradual shifts between solutions.
E) All of the above are true.

Question 17

Which is NOT true of learning vector quantization (LVQ) models?

A) LVQ is a type of artificial neural network methodology.
B) LVQ can handle both classification and regression problems.
C) LVQ requires supervised learning.
D) The researcher can set the learning rate of the LVQ algorithm.

Question 18

What is the Kappa statistic in the context of caret's machine learning models?

A) Kappa represents is the percent by which the researcher's model is better than random.
B) Kappa is one type of kernel for caret models.
C) Kappa is a visualization (plotting) parameter under caret.
D) Kappa is a tuning parameter under caret.

Question 19

The "mlr3" package is an alternative to which other widely-used R package?

A) rpart
B) kknn
C) caret
D) svm

Question 20

Which modeling package centers of creating and using "Learners" and "Tasks"? __________________.