Machine Learning With MATLAB Quick Reference

2/21/24, 5:59 PM Machine Learning with MATLAB - Quick Reference
Machine Learning with MATLAB
Getting Started
Review - Machine Learning Onramp
Summary: Machine Learning Onramp
These functions will be used throughout this course. Click

any function to visit the documentation.
View this information in the course quick reference guide

any time you need a reference.
Import and Process Data
allStats = readtable("bballStats.txt");
The readtable function creates a table
playerInfo = readtable("bballPlayers.txt");
in MATLAB from a data file.
positions = ["G","G-F","F-G","F","F-C","C-F","C"];
The categorical function creates a
playerInfo.pos = categorical(playerInfo.pos,positions);
categorical array from data.
allStats(19:end) = [];
Assigning the empty array removes
playerInfo = rmmissing(playerInfo);
rows or columns. The rmmissing
function removes any row with missing
or undefined elements.
playerStats = groupsummary(allStats,"playerID","sum");
The groupsummary function calculates
statistics grouped according to a
grouping variable.
data = innerjoin(playerInfo,playerStats);
The innerjoin function merges two
tables, retaining only the common key
variable observations.
Visualize and Engineer Features
boxplot(data.height,data.pos)
The boxplot function can create separate
box plots based on a grouping variable.
https://matlabacademy.mathworks.com/artifacts/quick-reference.html?course=mlml&language=en&release=R2023a 1/36
data{:,8:22} = data{:,8:22}./data.minutes;
You can use indexing and element-wise
data.minutes = data.minutes./data.GP;
division to scale variables in a table.
gscatter(data.rebounds,data.points,data.pos)
You can use the gscatter function to create

a grouped scatter plot.
Train and Evaluate a Model
knnmodel = fitcknn(data,"pos");
The fitcknn function fits a k-nearest
neighbors classification model.
knnmodel = fitcknn(data,"pos","NumNeighbors",5,...
You can use property name-value pairs to
"Standardize",true);
modify model options.
mdlLoss = loss(knnmodel,dataTest)
You can calculate the misclassification rate mdlLoss =
for a data set using the loss function. 0.4085
predPos = predict(knnmodel,dataTest);
The predict function uses a classification
model to predict classes for observations.
confusionchart(data.pos,predPos);
You can use the confusionchart function
to visually compare true classes and
predicted classes.
Finding Natural Patterns in Data

Low Dimensional Visualization
Multidimensional Scaling
Step 1 - Calculate pairwise distances
You can use the function pdist to calculate the pairwise distance between the observations. Note
that the input should be a numeric matrix.
>> D = pdist(data,"distance")
Outputs Inputs
D A distance or dissimilarity vector containing data An m-by-n numeric matrix
the distance between each pair of containing the data. Each of the
observations. D is of length m(m-1)/2. m rows is considered an
observation.
"distance" An optional input that indicates

the method of calculating the
distance or dissimilarity.
Commonly used methods are
"euclidean" (default),
"cityblock" , and
"correlation" .
Step 2 - Perform multidimensional scaling
You can now use the dissimilarity vector as an input to the function cmdscale .
>> [x,e] = cmdscale(D)
Outputs Inputs
x m-by-q matrix of the reconstructed D A distance or dissimilarity vector.
coordinates in q-dimensional space.
q is the minimum number of dimensions

needed to achieve the given pairwise
distances.
e Eigenvalues of the matrix x*x' .
You can use the eigenvalues e to determine if a low-dimensional approximation to the points in x
provides a reasonable representation of the data. If the first p eigenvalues are significantly larger than
the rest, the points are well approximated by the first p dimensions (that is, the first p columns of x ).
Principal Component Analysis
Another commonly used method for dimensionality reduction is principal component analysis (PCA).
Use the function pca to perform principal component analysis.
>> [pcs,scrs,~,~,pexp] = pca(data)
Outputs Inputs
pcs A n-by-n matrix of principal components. data An m-by-n numeric matrix. The n
columns correspond to n observed
scrs An m-by-n matrix containing the data
variables. Each of the m rows
transformed using the linear coordinate
corresponds to an observation.
transformation matrix pcs (first output).
pexp A vector of length n containing the

percentage of variance explained by
each principal component.
Interpreting the outputs
Suppose that the input matrix data has two columns which contain values of the observed variables x1 and x2.
[P,scrs,~,~,pexp] = pca(data)
P =
0.5687 0.8225
0.8225 -0.5687
After performing PCA, the first column of the output matrix P contains the coefficients of the first principal component.
[P,scrs,~,~,pexp] = pca(data)
P =
0.5687 0.8225
0.8225 -0.5687
The second column of the output matrix P contains the coefficients of the second principal component.
[P,scrs,~,~,pexp] = pca(data);
The second output scrs is a matrix containing the observations in data expressed in the coordinate space of the principal
components.
scrs(42,:)
ans =
1.0726 0.4163
For example, a single data point and its coordinates in the transformed space are shown here.
pexp
ans =
95.6706 4.3294
The last output pexp is a vector containing the percent variance explained by each principal component. Here, most of the
variance is explained by the first principal component.
scrs(:,1)
We can use only the first column of the transformed data, reducing the dimension of the data from 2 to 1.
Gaussian Mixture Models

GMM Clustering
Step 1 - Fit Gaussian Mixture Model
gm = fitgmdist(X,2);
You can use the function fitgmdist
to fit several multidimensional
gaussian (normal) distributions.
The command shown fits a mixture

gm of two distributions.
Step 2 - Identify Clusters
g = cluster(gm,X);
Now, the data can be clustered
probabilistically, by calculating each
observation’s posterior probability for
each component.
[g,~,p] = cluster(gm,X);
You can also return the individual
probabilities used to determine the
clusters.
The matrix p has two columns, one

for each of the two clusters.
Interpreting the Clusters

Visualizing Observations in Clusters
With high-dimensional data, it is difficult to visualize the groups as points in space. How can you
interpret the groups given by a clustering method?
Parallel Coordinates
Consider a data set in which each observation has 4 variables (measurements) x1, x2, x3, and x4.
Suppose that you have created two clusters.
You can visualize the first observation by plotting its variable value on the y-axis and the variable
number on the x-axis. Similarly, you can visualize the second observation. If the second observation
is in a different cluster, visualize it using a different color.
After visualizing several observations, you can see that cluster 1 has higher values of x1 and x3
whereas cluster 2 has higher values of x2 and x4.
Instead of visualizing each observation one-by-one, use the function parallelcoords , which creates
the graph shown above.
>> parallelcoords(X,"Group",g)
Inputs
X Data, specified as a numeric matrix.
"Group" Property name.
g A vector containing the observations'

group or cluster identifiers.
Evaluating Cluster Quality
When using clustering techniques such as k-means and Gaussian mixture models, you have to
specify the number of clusters. However, for high-dimensional data, it is difficult to determine the
optimum number of clusters.
You can use the silhouette values to judge the quality of the clusters. An observation’s silhouette
value is a normalized measure (between -1 and +1) of how close that observation is to other
observations in the same cluster, compared to the observations in different clusters.
Silhouette Plots
A silhouette plot shows the silhouette value of each observation, grouped by cluster. Clustering
schemes in which most of the observations have high silhouette value are desirable.
Use the silhouette function in MATLAB to create a silhouette plot.
Below are two silhouette plots for the same data set. On the left, the observations are divided into two
clusters, and on the right they are divided into three clusters.
[grp,c] = kmeans(X,2); [grp,c] = kmeans(X,3);

silhouette(X,grp) silhouette(X,grp)
In this case, dividing the data into 2 clusters instead of 3 produces better quality clusters. The
silhouette plot of two clusters shows fewer negative silhouette values, and those negative values are
of smaller magnitude than the negative values in the silhouette plot of three clusters.
Automate Cluster Quality Evaluation
Instead of manually experimenting with different numbers of clusters, you can automate the process
with the evalclusters function.
The following function call creates 2, 3, 4, and 5 clusters using k-means clustering, and calculates the
silhouette value for each clustering scheme.
clustev = evalclusters(X,"kmeans","silhouette","KList",2:5)
The output variable, clustev , contains detailed information about the evaluation including the
optimum number of clusters.
kbest = clustev.OptimalK
In place of "silhouette" , you can use other evaluation criteria such as "CalinskiHarabasz" ,
"DaviesBouldin" , and "gap" . Refer to the documentation for further details.
Hierarchical Clustering
Hierarchical Clustering
Step 1 – Determine Hierarchical Structure
Finding the hierarchical structure involves calculating the distance between each pair of points and
then using these distances to link together pairs of “neighboring” points.
Z = linkage(X,"ward","cosine");
Use the linkage function to create the
hierarchical tree.
The optional second and third inputs specify

the methods for calculating the distance
between clusters (default: "single" ) and
calculating the distance between points
(default: "euclidean" ).
dendrogram(Z)
You can use the dendrogram function to

visualize the hierarchy.
Step 2 – Divide Hierarchical Tree into Clusters
You can use the cluster function to assign observations into groups, according to the linkage
distances Z .
Z = linkage(X,"centroid","cosine");
dendrogram(Z)
grp = cluster(Z,"maxclust",3)
Classification Methods
Nearest Neighbor Classification
k-Nearest Neighbor Overview
Function fitcknn
Fit Time Prediction Time Memory Overhead

Performance Fast Fast Small
∝ (Data Size)2
"NumNeighbors" – Number of neighbors used for classification. (Default: 1)

Common
"Distance" – Metric used for calculating distances between neighbors.
Properties
"DistanceWeight" – Weighting given to different neighbors.
For normalizing the data, use the "Standardize" option.

Special
Notes The "cosine" distance metric works well for “wide” data (more predictors
than observations) and data with many predictors.
Classification Trees
Decision Trees Overview
Function fitctree

Performance
∝ Size of the data Fast Small
"SplitCriterion" – Formula used to determine optimal splits at each level.

Common "MinLeafSize" – Minimum number of observations in each leaf node.
Properties "MaxNumSplits" – Maximum number of splits allowed in the decision tree.
Special
Trees are a good choice when there is a significant amount of missing data.
Notes
Naive Bayes Classification

Naïve Bayes Overview
Function fitcnb

Normal Dist. - Fast Normal Dist. - Fast Normal Dist. - Small
Performance
Kernel Dist. - Slow Kernel Dist. - Slow Kernel Dist. -
Moderate to large
"DistributionNames" – Distribution used to calculate probabilities.

Common "Width" – Width of the smoothing window (when "DistributionNames" is set to
Properties "kernel" ).
"Kernel" – Type of kernel to use (when "DistributionNames" is set to "kernel" ).
Special
Naive Bayes is a good choice when there is a significant amount of missing data.
Notes
Discriminant Analysis
Fitting Discriminant Analysis Models
Linear Discriminant Analysis

By default, the covariance for each response class is assumed to be the same. This results in linear
boundaries between classes.
daModel = fitcdiscr(dataTrain,"ResponseVarName")
Quadratic Discriminant Analysis

Removing the assumption of equal covariances results in a quadratic boundary between classes. Use
the "DiscrimType" option to do this:
daModel = fitcdiscr(dataTrain,"ResponseVarName","DiscrimType","quadratic")
Discriminant Analysis (DA) Review
Function fitcdiscr

Fast Fast Linear DA - Small
Performance ∝ size of the data ∝ size of the data Quadratic DA -
Moderate to large
∝ number of predictors
"DiscrimType" - Type of boundary used.

Common "Delta" - Coefficient threshold for including predictors in a linear boundary. (Default: 0)
Properties
"Gamma" - Regularization to use when estimating the covariance matrix for linear DA.
Special Linear discriminant analysis works well for “wide” data (more predictors than
Notes observations).
Support Vector Machines

Support Vector Machines Review
Function fitcsvm

Fast Very Fast Moderate
Performance
∝ square of the size of ∝ square of the size of
the data the data
"KernelFunction" – Variable transformation to apply.

Common "KernelScale" – Scaling applied before the kernel transformation. (Default value: 1)
Properties "BoxConstraint" – Regularization parameter controlling the misclassification penalty. (Default
value: 1)
SVMs use a distance based algorithm. For data that is not normalized, use the "Standardize"
Special option.
Notes
Linear SVMs work well for “wide” data (more predictors than observations). Gaussian SVMs often
work better on “tall” data (more observations than predictors).
Multiclass Support Vector Machine Models
The underlying calculations for classification with SVMs are binary by nature. You can perform
multiclass SVM classification by creating an error-correcting output codes (ECOC) classifier.
Suppose there are three classes in the data.
By default, the ECOC model reduces the model to multiple, binary classifiers using the one-vs-one
design.
The combination of the resulting classifiers is used for prediction.
Creating Multiclass Support Vector Machines
Creating a multiclass SVM model is a two-step process.
1. Create a template for a binary classifier – Create a template for a binary SVM using the function templateSVM .
>> template = templateSVM("PropertyName",PropertyValue)
Outputs Inputs
template An SVM classifier. "PropertyName" Optional property
name, e.g.,
"KernelFunction" .
PropertyValue Property value, e.g.,

"polynomial" .
You can provide

multiple property
name-value pairs.
2. Create multiclass SVM classifier – Use the function fitcecoc to create a multiclass SVM classifier.
>> ecocModel = fitcecoc(dataTrain,"y","Learners",template)
Outputs Inputs
ecocModel ECOC classifier. dataTrain Training data.
"y" Response variable

name.
"Learners" Property name for

specifying the binary
classifier.
template Binary classifier.
Classification with Neural Networks

Neural Networks
Function fitcnet

Slow Slow Medium
Performance
∝ Network & data size ∝ Network & data size ∝ Network & data size
"LayerSizes" – Sizes of fully connected layers. (Default: 10)

Common "Activations" – Activation functions for fully connected layers. (Default: ReLU)
Properties "ValidationData" – Validation data for training convergence detection.
Special
Neural networks require data to be normalized. For normalizing the data, use the
Notes
"Standardize" option.
To train a neural network with more than one hidden layer specify the number of neurons
per hidden layer as a vector. For example, "LayerSizes",[20,15] creates a network
with two hidden layers with 20 and 15 neurons accordingly.
Neural networks work well for "tall" data (more observations than predictors).
Improving Predictive Models
Cross Validation
Using Cross Validation
When you create a model, the model variable

contains information about both the mathematical
model and the data set.
The information about the partitions for cross

validation will also be stored in the model variable.
Fitting Cross-Validated Models
To create a model with cross validation, provide one of the following options in the model-creation
function.
mdl = fitcknn(data,"Response","PropertyName",PropertyValue)
Property Name Property Value Type of validation
"CrossVal" "on" 10-fold cross validation
"Holdout" scalar from 0 to 1 Holdout with the given fraction reserved for validation.
"KFold" k (scalar) k-fold cross validation
"Leaveout" "on" Leave-one-out cross validation
If you already have a partition created using the cvpartition function, you can provide that to the
fitting function instead.
cvpt = cvpartition(data.Response,"KFold",k)
mdl = fitcknn(data,"Response","CVPartition",cvpt)
Calculating the Loss
To evaluate a cross-validated model, use the kfoldLoss function.
mdlLoss = kfoldLoss(mdl)
Reducing Predictors - Feature Transformation

Principal Component Analysis
Principal component analysis (PCA) transforms an n-dimensional feature space into a new n-
dimensional space of orthogonal components. The components are ordered by the variation
explained in the data.
>> [pcs,scrs,~,~,pctExp] = pca(X)
Outputs Inputs
pcs Principal components X Data matrix with m observations and
coefficients. The contribution of n predictors.
each predictor to the n principal (m-by-n matrix)
components.
(n-by-n matrix)
scrs Principal component scores.

The representations of the m
observations in X in the
principal component space.
(m-by-n matrix)
pctExp Percent of the variance in the

data explained by each
principal component.
(1-by-n vector)
The transformed variables contain the same amount of information as the original data. However,
assuming that the data contains some amount of noise, the components that contain the last few
percent of explained variance are likely to represent noise more than information.
Reducing the Number of Predictors
PCA can be used for dimensionality reduction by discarding the principal components beyond a
chosen threshold of explained variance. The pareto function can be used to visualize the variance
explained by the principal components.
In the following example, the input X has 11 columns but the first 9 principal components explain
more than 95% of variance.
[pcs,scrs,~,~,pctExp] = pca(X);
pareto(pctExp)
Xreduced = scrs(:,1:9);
Interpreting Principal Components
The principal components by themselves have no physical meaning. However, the coefficients of the
linear transformation indicate the contribution of each variable in the principal component.
For example, if the coefficients of the first principal component are 0.8, 0.05, and 0.3, the first variable
has the largest contribution followed by the third and the second variable.
Biplot
You can visualize any two principal components using the function biplot . It's commonly used to
visualize the first two principal components, which explain the greatest amount of variance in the
data.
In the following biplot, you can see that the predictive variables Age , InducedSTDep , METS , and
ExerciseDuration contribute heavily to the first principal component, but not to the second principal
component.
biplot(pcs(:,1:2),"VarLabels",varnames)
Heat Map
You can also visualize the contributions of each variable to the principal components as a heat map.
For example, the following heat map visualizes the first three principal components. You can see that
METS , MaxHeartRate , and ExerciseDuration contribute heavily to the first principal component.
heatmap(abs(pcs(:,1:3)),...
"YDisplayLabels",varnames);
xlabel("Principal Component")
Parallel Coordinates Plot
PCA can be performed independent of the response variable. However, when the data has a
response variable that has multiple categories (for example, true and false), a parallel coordinates
plot of the principal component scores can be useful.
In the following plot, notice that the observations from one group (false) have high values of the first
principal component and the observations from the second group (true) have low values.
parallelcoords(scrs,"Group",y,"Quantile",0.25)
Reducing Predictors - Feature Selection

Feature Ranking Algorithms
Feature ranking algorithms assign scores to features based on how relevant they are to the model
according to a given metric. Some common algorithms include Chi-Square, Minimum Redundancy
Maximum Relevance (MRMR), and Neighborhood Component Analysis.
Use the fs__ functions to apply feature ranking algorithms to data. Most feature ranking algorithms
return the positions and scores for all the features sorted from highest to lowest.
>> [idx,scores] = fs__(tblData,ResponseVarName)
Outputs Inputs
idx Indices of predictors ordered by tblData Table data with predictor
predictor importance. variables and a response
variable.
scores Predictor scores.
ResponseVarName Name of the response
variable.
You can then use linear indexing to select a subset of features and use them to train the model.
toKeep = idx(1:nFeatures);
selected = [tblData(:,toKeep),tblData(:,"ResponseVarName")];
mdl = fitcknn(selected,"ResponseVarName");
Feature ranking algorithms in MATLAB
Algorithm Function Task
Chi-Square fscchi2 Classification
Minimum redundancy maximum relevance (MRMR) fscmrmr / fsrmrmr Classification/Regression
Neighborhood component analysis (NCA) fscnca / fsrnca Classification/Regression
F-tests fsrftest Regression
Laplacian fsulaplacian Clustering
ReliefF/RReliefF relieff Classification/Regression
Sequential Feature Selection
The function sequentialfs performs sequential feature selection.
>> toKeep = sequentialfs(fun,X,y)
Outputs Inputs
toKeep Logical vector indicating which fun Function handle for a function that
predictors are included in the fits a model and calculates the loss.
final model.
X Numeric matrix with m observations
and n predictors.
y Numeric vector of m response

values.
Note that the first input is a handle to the error function.
tokeep = sequentialfs(@errorFun,X,y)
You can use the optional property "cv" to specify the cross validation method. For example, you can
specify a 7-fold cross validation.
tokeep = sequentialfs(@errorFun,X,y,"cv",7)
Structure of the Error Function
Sequential feature selection requires an error function that builds a model and calculates the
prediction error. The error function must have the following structure.
Four numeric inputs:
two for training the model (predictor matrix and response vector)
two for evaluating the model (predictor matrix and response vector)
One scalar output representing the prediction error.
errorFun.m
1. function error = errorFun(Xtrain,ytrain,Xtest,ytest)
2.
3. % Create the model with the learning method of your choice
4. mdl = fitcsvm(Xtrain,ytrain);
5.
6. % Calculate the number of test observations misclassified
7. ypred = predict(mdl,XTest);
8. error = nnz(ypred ~= ytest);
9.
10. end
Note You do not need to create the training and test data sets. The sequentialfs function will
internally partition the data before calling the error function.
Accommodating Categorical Data
Some algorithms and functionality (e.g., sequentialfs ) requires predictors in the form of a numeric
matrix. If your data contains categorical predictors, how can you include these predictors in the
model?
One option is to assign a number to each category. However, this may impose a false numerical
structure on the observations.
For example, say you assign the numbers 1 through 4 to four categories in a predictor. This implies
that the distance between categories 1 and 4 is three times larger than the distance between
categories 3 and 4. In reality, the categories are probably equidistant from each other.
Creating Dummy Variables
A better approach is to create new dummy predictors, with one dummy

predictor for each category.
You can create a matrix of dummy variables using the function dummyvar .
d = dummyvar(c)
Each column of d represents one of the categories in c . Each row

corresponds to an observation, and has one element with value 1 and all
other elements equal to 0 . The 1 appears in the column corresponding to
that observation's assigned category.
This matrix can now be used in a machine learning model in place of the categorical vector c , with
each column being treated as a separate predictor variable that indicates the presence ( 1 ) or
absence ( 0 ) of that category of c .
Concatenate this matrix with a matrix containing the numeric predictors.
Hyperparameter Optimization
Performing Hyperparameter Optimization
You can use the "OptimizeHyperparameters" property name-value pair to choose which model properties to optimize.
Most model-creation functions accept the "OptimizeHyperparameters" option.
>> mdl = fitcknn(data,"ResponseName","OptimizeHyperparameters",params)
Outputs Inputs
mdl Model fit using data,"ResponseName" Table of predictors and
optimized property values. response values and
response variable name.
"OptimizeHyperparameters" Optional property for

hyperparameter
optimization.
params Model properties to optimize,

specified as a string array or
cell array. Using "auto" and
"all" optimizes pre-
selected properties.
During the optimization, iterative updates are displayed, as well as a plot with the best objective function value against the
iteration number.
mdl = fitcknn(data,"y","OptimizeHyperparameters","auto")
Setting the "OptimizeHyperparameters" property value to "auto" will optimize a typical set of hyperparameters. The
properties optimized differ depending on the model type. For example, for Nearest Neighbor classification, the optimized
properties are "Distance" and "NumNeighbors" .
Optimization Options
By default, hyperparameter optimization uses Bayesian optimization and tries to minimize the 5-fold cross-validation loss. You
can change these settings with the "HyperparameterOptimizationOptions" property name-value pair.
Specify the optimization options using a structure. To use 10-fold cross validation, create a cross-validation partition and then
create a structure containing option name-value pairs.
part = cvpartition(y,"KFold",10);
opt = struct("CVPartition",part);
mdl = fitcknn(data,"y","OptimizeHyperparameters","auto","HyperparameterOptimizationOptions",opt);
You can set many optimization options in the structure. For example, you can hide the plots and set the maximum number of
objective function evaluations.
opt = struct("ShowPlots",false,"MaxObjectiveEvaluations",50);
To see the available options for a particular model-creation function, view the function's documentation.
Ensemble Learning
Fitting Ensemble Models
The fitcensemble function creates a classification ensemble of weak learners. Similarly, the fitrensemble function
creates a regression ensemble. Both functions have identical syntax.
>> mdl = fitcensemble(data,"ResponseName")
Outputs Inputs
mdl Ensemble model variable. data Table containing the
predictors and
response values.
"ResponseName" Response variable

name.
Commonly Used Options
"Method" - Bagging (bootstrap aggregation) and boosting are the two most common approaches used in ensemble
modeling. The fitcensemble function provides several bagging and boosting methods. For example, use the "Bag"
method to create a random forest.
mdl - fitcensemble(data,"Y","Method","Bag")
The default method depends on if it is a binary or multiclass classification problem, as well as the type of learners in the
ensemble.
"Learners" - You can specify the type of weak-learner to use in the ensemble: "tree" , "discriminant" , or "knn" .
The default learner type depends on the method specified: the method "Subspace" has default learner "knn" , and all
other methods have default learner "tree" .
mdl = fitcensemble(data,"Y","Learners","knn")
The fitcensemble function uses the default settings for each learner type. To customize learner properties, use a weak-
learner template.
mdl = fitcensemble(data,"Y","Learners",templateKNN("NumNeighbors",3))
You can use a cell vector of learners to create an ensemble composed of more than one type of learner. For example, an
ensemble could consist of two types of kNN learners.
lnrs = {templateKNN("NumNeighbors",3),templateKNN("NumNeighbors",5)}
mdl = fitcensemble(data,"Y","Learners",lnrs)
"NumLearningCycles" - At every learning cycle, one weak learner is trained for each learner specified in "Learners" .
The default number of learning cycles is 100. If "Learners" contains only one learner (as is usually the case), then by
default 100 learners are trained. If "Learners" contains two learners, then by default 200 learners are trained (two
learners per learning cycle).
Regression Methods
Linear Models
Fitting Linear Regression Models
Use the function fitlm to fit a linear regression model.
>> mdl = fitlm(data,"modelspec")
Outputs Inputs
mdl A regression model variable containing the data A table containing the data used
coefficients and other information about to fit the regression model. See
the model. below for details.
"modelspec" Specification of the regression

model. See below for details.
How to Organize the Data
In a regression model, the relationship between the The first input to fitlm is a table containing the predictors
predictors and the response can be described by the and the response. By default, fitlm uses the last column as
following formula. the response and all other columns as predictors.
How to Specify the Regression Model
When modeling a linear regression, you can apply different functions to the predictive variables. As the second input to
fitlm , you can use one of the predefined models or you can specify a model by providing a formula in Wilkinson–Rogers
notation.
Common predefined models: Wilkinson–Rogers notation: responseVar ~ terms
Model name Meaning Operator Meaning Example
"linear" Intercept and linear terms + Include "y ~ x1+x2" includes the intercept
for each predictor. this term. term, x1, and x2:
y = c0 + c1 x1 + c2 x2
"interactions" Intercept, linear terms, and
all products of pairs of - Exclude "y ~ x1+x2-1" excludes the intercept
distinct predictors (no this term. term:
squared terms). y = c1 x1 + c2 x2
"quadratic" Intercept, linear terms,
* Include "y ~ x1*x2" includes the intercept
interactions, and squared
product term, x1, x2, and x1*x2:
terms.
and all y = c0 + c1 x1 + c2 x2 + c3 x1 x2
lower-
order
terms.
: Includes "y ~ x1:x2" includes only x1*x2:

the y = c1 x1 x2
product
term only.
^ Include "y ~ (x1^2)+(x2^2)" includes the

power intercept, x1, x2, x1^2, and x2^2:
and all y = c0 + c1 x1 + c2 x2 + c3 x12 + c4 x22
lower-
Operator Meaning Example

order
terms.
Fitting Models Using Data Stored in Arrays
If your data is stored in numeric arrays instead of a single table, you can provide the fitlm function
the predictor and response values as numeric arrays.
>> mdl = fitlm(X,y)
Outputs Inputs
mdl Linear regression model X Predictor values, specified as a
variable. numeric matrix.
y Response values, specified as a

vector.
How to Specify the Regression Model
Each column of the predictor matrix X is treated as one predictor variable. By default, fitlm will fit a
model with an intercept and a linear term for each predictor (column).
To fit a different regression formula, you have two options. You can store the predictors and the
response in a table and provide the model specification separately. Alternatively, you can create a
matrix with a column for each term in the regression formula. This matrix is called a design matrix.
Stepwise Fitting
Fitting Stepwise Linear Regression
Use the function stepwiselm to fit a stepwise linear regression model.
>> stepwiseMdl = stepwiselm(data,"modelspec")
Outputs Inputs
stepwiseMdl A linear model variable data A table containing the
containing the coefficients data used to fit the
and other information regression model. See
about the model. below for details.
"modelspec" Starting model for the

stepwise regression (see
below).
As with fitlm , stepwiselm uses the last column of data as the response and all other columns as predictors.
How to Specify a Model
stepwiselm chooses the model for you. However, you can provide the following inputs to control the
model selection process.
"modelspec" - The second input to the function specifies the starting model. stepwiselm starts
with this model and adds or removes terms based on certain criteria.
Commonly used starting values: "constant" , "linear" , "interactions" (constant, linear,

and interaction terms), "quadratic" (constant, linear, interaction, and quadratic terms).
"Lower" and "Upper" - If you want to limit the complexity of the model, use these properties.
For example, the following model will definitely contain the intercept and the linear terms but will
not contain any terms with a degree of three or more.
mdl = stepwiselm(data,"Lower","linear","Upper","quadratic")
By default, stepwiselm considers models as simple as a constant term only and as complex as
an interaction model.
How a Model is Chosen
stepwiselm iteratively adds and subtracts terms from the starting model, if the modified model is
better than the previous iteration.
“Better” is judged according to the value of the "Criterion" property. The default value is "sse" —
an F-test on the sum of squared error. You can change this measure by setting the "Criterion"
property.
mdl = stepwiselm(data,"Criterion","rsquared")
Available values: "sse" (default), "aic" , "bic" , "rsquared" , "adjrsquared"
Regularized Linear Models

Ridge and Lasso Regression
In linear regression, the coefficients are chosen by minimizing the mean squared error (MSE). Mean
squared error is the squared difference between the observed and the predicted response value.
In ridge and lasso regression, a penalty term is added to MSE.
This penalty term is composed of the fit coefficient values and a tuning parameter λ. The larger the
value of λ, the greater the penalty and, therefore, the more the coefficients are “shrunk” towards zero.
The Penalty Term
The difference between the two methods is how the penalty term is calculated. Ridge regression uses
an L2 norm of the coefficients. Lasso uses an L1 norm.
The use of different norms provides different regularization behavior.
Ridge regression shrinks coefficients continuously, and keeps all predictors.

Lasso allows coefficients to be set to zero, reducing the number of predictors included in the
model.
Lasso can be used as a form of feature selection, however feature selection may not be appropriate
for cases with similar, highly correlated variables. It may result in loss of information which could
impact accuracy and the interpretation of results. Ridge regression maintains all features, but the
model may still be very complex if there is a large number of predictors.
Elastic Net
You can also use a penalty term that uses a weighted average of both. This is elastic net regression,
which introduces another parameter – the weighting between ridge (L2 norm) and lasso (L1 norm).
Fitting Ridge Regression Models
You can use the function ridge to fit a ridge regression model.
>> b = ridge(y,X,lambda,scaled)
Outputs Inputs
b Ridge regression y Response values,
coefficients. specified as a vector.
X Predictor values,
specified as a
numeric matrix.
lambda Regularization
parameter.
scaled A {0,1}-valued flag to

determine if the
coefficient estimates
in b are restored to
the scale of the
original data.
Notes
The matrix X is a numeric design matrix with columns representing the terms in the regression
formula. If the original data contains two predictive variables x1 and x2, but the desired regression
model formula contains the terms x1, x2, and x1*x2, the matrix X should have 3 columns: x1, x2, and
x1*x2.
The ridge parameter lambda is a non-negative number. In a later section, you will try to estimate the
optimum value of lambda .
The ridge function normalizes the predictors before fitting the model. Therefore, by default, the
regression coefficients correspond to the normalized data. Set the scaled flag to 0 to restore the
coefficients to the scale of the original data.
Predicting Response using Ridge Regression Models
When using a ridge regression model for predictions, you will want the regression coefficients in the
scale of the original data. In this case (the scaled flag set to 0 ), the coefficient vector b will contain
n+1 coefficients for a model with n predictors. The first element of b corresponds with the intercept
term.
You can predict the response by multiplying the matrix containing the predictors and the last n
elements of the coefficient vector. Add the first element of the coefficient vector to incorporate the
intercept in the calculation.
yPred = dataNew*b(2:end) + b(1);
Determining the Regression Parameter
When modeling data using ridge regression, how can you determine the regression parameter,
lambda ?
When you provide a vector of λ values to the ridge

function, the output b is a matrix of coefficients.
b = ridge(y,X,lambda,scaled);
The columns of b contain the coefficient values for each

parameter in the vector lambda .
You can now use each column of the matrix b as regression coefficients and predict the response.
yPred = dataNew*b(2:end,:) + b(1,:);
The response yPred is a matrix where each column is the predicted response for the corresponding
value of lambda . You can use yPred to calculate the mean squared error (MSE), and choose the
coefficients which minimize MSE.
Fitting Lasso Regression Models
Lasso (Least Absolute Shrinkage and Selection Operator) regression models are fit using the lasso
function.
>> [b,fitInfo] = lasso(X,y,"Lambda",lambda)
Outputs Inputs
b Lasso coefficients. X Predictor values,
specified as a
fitInfo A structure
numeric matrix.
containing
information about y Response values,
the model. specified as a
vector.
"Lambda" Property name for

specifying the
regularization
parameter.
lambda Regularization
parameter value.
Notes
Like in ridge, the matrix X is a design matrix with columns representing the terms in the regression
formula.
The "Lambda" property is optional. If not specified, lasso uses a geometric sequence of
λ values based on the data.
ridge and lasso implement their penalty terms slightly differently, and as a result, use different
scalings for lambda . To use λ values in lasso which have the same interpretation as for ridge ,
scale lambda in lasso by the number of observations.
Use the optional property "Alpha" with a value between 0 and 1 to create an elastic net. Recall that
elastic net regression uses a penalty term which is a weighted average of the ridge (L2) and lasso
(L1) penalty terms. "Alpha" values near 1 are closer to lasso, and "Alpha" values near 0 are closer
to ridge.
Predicting Response using Lasso Model
You can predict the response by multiplying the matrix containing the predictors by the coefficient
vector.
Note that the intercept term is not included in the output coefficients. Instead, it is a field in the output
structure fitInfo .
yPred = dataNew*b + fitInfo.Intercept
SVMs, Trees and Neural Networks

Regression using SVMs, Decision Trees, and Neural Networks
Linear regression techniques like fitlm are parametric, which means predictions are
based on a finite set of parameters that are estimated from the training data.
A parametric regression model:

Assumes a relationship that can be specified using a formula.
Is easy to interpret – you can measure the change in the response per unit
change in the predictor.
You may not need a model with a specific interpretable formula if its primary purpose
is simply to predict the response for unknown observations. In this case, you can use
a nonparametric model.
A nonparametric regression model:

Does not fit the regression model based on a given formula.
Can provide more accurate predictions, but is more difficult to interpret.
Support vector machines (SVMs), decision trees, and neural networks are some of the nonparametric
techniques you can use for regression.
fitrtree Fit binary regression decision tree.
fitrsvm Fit a support vector machine regression model.
fitrnet Fit a neural network for regression.
These techniques are covered in detail in the Classification Methods chapter. You can find information
specific to regression in the documentation.
Gaussian Process Regression

Fitting and Predicting Using Gaussian Process Regression
Fitting GPR Models
You can fit a GPR model by using the fitrgp function.
>> mdl = fitrgp(data,"ResponseVarName")
Outputs Inputs
mdl A GPR model variable. data A table containing
the predictor and
response values.
"ResponseVarName" Name of the

response variable.
Use the "KernelFunction" property to change the kernel to one of the predefined options.
mdl = fitrgp(data,"ResponseVarName","KernelFunction","exponential")
Predicting the Response
In addition to the predicted response values, the predict function for GPR models can also return
the standard deviation and prediction intervals for the predicted values.
>> [yPred,yStd,yInt] = predict(mdl,dataNew)
Outputs Inputs
yPred Predicted response value(s). mdl A GPR model variable.
yStd Standard deviation for each predicted dataNew Predictor values for one or more new
value. observations.
yInt Matrix whose columns contain the lower

and upper limits of the 95% prediction
interval for each predicted value.
You can change the significance level of the prediction intervals by setting the "Alpha" property to a
value between 0 and 1. The default value is 0.05.
[yPred,yStd,yInt] = predict(mdl,dataNew,"Alpha",0.01)

Machine Learning With MATLAB Quick Reference

Uploaded by

Machine Learning With MATLAB Quick Reference

Uploaded by

2/21/24, 5:59 PM Machine Learning with MATLAB - Quick Reference

Machine Learning with MATLAB

These functions will be used throughout this course. Click

View this information in the course quick reference guide

Import and Process Data

Visualize and Engineer Features

You can use the gscatter function to create

Train and Evaluate a Model

Finding Natural Patterns in Data

Step 1 - Calculate pairwise distances

"distance" An optional input that indicates

Step 2 - Perform multidimensional scaling

>> [x,e] = cmdscale(D)

q is the minimum number of dimensions

e Eigenvalues of the matrix x*x' .

Principal Component Analysis

>> [pcs,scrs,~,~,pexp] = pca(data)

pexp A vector of length n containing the

Interpreting the outputs

Gaussian Mixture Models

Step 1 - Fit Gaussian Mixture Model

The command shown fits a mixture

Step 2 - Identify Clusters

The matrix p has two columns, one

Interpreting the Clusters

"Group" Property name.

g A vector containing the observations'

Evaluating Cluster Quality

Use the silhouette function in MATLAB to create a silhouette plot.

[grp,c] = kmeans(X,2); [grp,c] = kmeans(X,3);

Automate Cluster Quality Evaluation

Step 1 – Determine Hierarchical Structure

The optional second and third inputs specify

You can use the dendrogram function to

Step 2 – Divide Hierarchical Tree into Clusters

Fit Time Prediction Time Memory Overhead

"NumNeighbors" – Number of neighbors used for classification. (Default: 1)

For normalizing the data, use the "Standardize" option.

Fit Time Prediction Time Memory Overhead

"SplitCriterion" – Formula used to determine optimal splits at each level.

Naive Bayes Classification

Fit Time Prediction Time Memory Overhead

"DistributionNames" – Distribution used to calculate probabilities.

Linear Discriminant Analysis

Quadratic Discriminant Analysis

Discriminant Analysis (DA) Review

Fit Time Prediction Time Memory Overhead

"DiscrimType" - Type of boundary used.

Support Vector Machines

Fit Time Prediction Time Memory Overhead

"KernelFunction" – Variable transformation to apply.

Multiclass Support Vector Machine Models

Suppose there are three classes in the data.

The combination of the resulting classifiers is used for prediction.

Creating Multiclass Support Vector Machines

Creating a multiclass SVM model is a two-step process.

>> template = templateSVM("PropertyName",PropertyValue)

PropertyValue Property value, e.g.,

You can provide

>> ecocModel = fitcecoc(dataTrain,"y","Learners",template)

"y" Response variable

"Learners" Property name for

template Binary classifier.

Classification with Neural Networks

Fit Time Prediction Time Memory Overhead

"LayerSizes" – Sizes of fully connected layers. (Default: 10)

Improving Predictive Models

When you create a model, the model variable

The information about the partitions for cross

Fitting Cross-Validated Models

Property Name Property Value Type of validation

"CrossVal" "on" 10-fold cross validation

"KFold" k (scalar) k-fold cross validation

"Leaveout" "on" Leave-one-out cross validation

Calculating the Loss

To evaluate a cross-validated model, use the kfoldLoss function.