Question 1

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

Accepted Answer

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

Accepted Answer

The support vector machine (SVM) method is a powerful tool for solving classification and regression problems. Therefore, it can be applied to both B (classification problems) and C (regression problems). It is not suitable for addressing assembly line process problems, which require a different approach.

Question 3

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

Accepted Answer

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

Accepted Answer

The probabilities generated by SVM are not based on ordinary statistical inference but rather on cross-validation.

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.

Accepted Answer

All the options listed (A, B, and C) are considered negative aspects of SVM (Support Vector Machines) methods. SVM can be seen as less transparent or a "black box" compared to simpler models like OLS regression, optimizing SVM models often requires trial and error to select the right kernel and parameters, and despite using techniques like cross-validation, it's still possible to overfit SVM models to noise in the data.

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

Accepted Answer

Margins: SVM tries to maximize the separation, called the "margin", between classes being predicted. Hyperplanes classify the observed data points into distinct classes but their number is not what is optimized. Support vectors are the data points which are closest to the margin. The number of support vectors may be very small even if the number of observations is very large, and their number is not maximized. Cost is a parameter which may be set by the researcher. Too high a cost and the algorithm will work toward 100% separation but at the price of a complicated hyperplane which may not generalize well to other data in the future. Such a model has narrow margins and is "overfitted", meaning it is responsive even to noise in the data. If cost is too low then the margin will be wide, allowing more error in the model. Cost is not minimized. Percent variance explained is the metric for OLS regression, not SVM.

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:

Accepted Answer

The answer of A "loss function" is a metric to...

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

Accepted Answer

Kernels are used in SVM to transform the input data into a higher-dimensional space, where a linear separation of the data is possible. This is particularly useful when dealing with nonlinearly separable data. Kernels function as a similarity function between points in this higher-dimensional space, allowing the SVM to find the optimal hyperplane that separates the data points. Therefore, A is the correct choice. B, C, and D are not accurate descriptions of the purpose of kernels in 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

Accepted Answer

Also called the radial basis function (RBF) or Gaussian kernel, this type of kernel has similarities to the algorithm for RBF neural networks. This is the default kernel in the svm command of the e1071 package. With a radial kernel the researcher is assuming that data may be separated better by circles and hyperspheres than by straight lines and hyperplanes.

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

Accepted Answer

Gamma, degree, coef0, and nu are all tuning parameters in SVM that can be optimized to achieve the best performing model. They are used to adjust the complexity of the decision boundary and to control the trade-off between fitting the training data well and generalizing to new data. Coefficients are not labels, and gamma, degree, coef0, and nu do not refer to stages in the problem-solving process.

Question 11

SVM routinely outperforms OLS regression when the data a nonlinear?

Accepted Answer

SVM is a nonlinear model and can accommodate nonlinear relationships between variables, making it better suited for nonlinear data than OLS regression, which assumes a linear relationship between variables.

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

Accepted Answer

Variable importance may be assessed through a leave-one-out method in which one predictor at a time is dropped from the model.

Question 13

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

Accepted Answer

In online help documentation, David Meyer, developer of the svm program, warns, "Parameters of SVM-models usually must be tuned to yield 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.

Accepted Answer

The train command of the caret package allows for tuning of machine learning models based on built-in defaults and user-specified parameters. This allows for the selection of optimal hyperparameters for the model being trained.

Question 15

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

Accepted Answer

This is true of OLS but not of SVM. In fact, the classic use of SVM was for classification of binary outcome variables, though it can handle both categorical and continuous response variables.

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.

Accepted Answer

All the statements provided are true regarding gradient boosting machine (GBM) models. GBM can indeed be implemented under the "caret" package in R, which is a comprehensive framework for building machine learning models. Both decision trees and random forests can serve as base learners in GBM, as GBM is a flexible algorithm that can utilize various types of models as its base learners, including decision trees which are the most common. OLS (Ordinary Least Squares) regression can also be used as a base learner in GBM, especially for regression tasks. Lastly, the iterative process of GBM involves weighting cases differentially to gradually improve the model's accuracy, which correctly describes how GBM optimizes its solution.

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.

Accepted Answer

LVQ is for classification and not regression problems. It can classify binary and categorical outcome variables with multiple levels. "Supervised" means that the researcher must input a set of training patterns (data rows) for which the value of the output variable to be classified is known. The amount of the adjustment in a given iteration is determined by the "learning rate", which the researcher may set.

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.

Accepted Answer

The Kappa statistic in caret's machine learning models is a measure of inter-rater agreement between the predicted and actual classes. It represents the percentage by which the researcher's model is better than random, and is often used to evaluate the performance of classification models.

Question 19

The "mlr3" package is an alternative to which other widely-used R package?
A) rpart
B) kknn
C) caret
D) svm

Accepted Answer

The mlr3 package is a newer alternative to caret. Both are modeling integration packages supporting a variety of specific modeling and machine learning methods. The mlr3 package, for instance, supports A, B, and D.

Question 20

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

Accepted Answer

The answer of Which modeling package centers of creating and...

Quiz 6: Modeling and Machine Learning

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

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

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

Which is NOT a positive aspect of SVM methods?

Which is NOT a negative aspect of SVM methods?

What does the SVM algorithm attempt to optimize?

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:

What is the purpose of "kernels" in SVM?

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

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

SVM routinely outperforms OLS regression when the data a nonlinear?

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

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

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.

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

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

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

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

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

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