Multiclass problems

Support Vector Machines in R

Kailash Awati

Instructor

The iris dataset - an introduction

150 measurements of 5 attributes
- Petal width and length - number (predictor variables)
- Sepal width and length - number (predictor variables)
- Species - category: setosa, virginica or versicolor (predicted variable)
Dataset available from UCI ML repository

Visualizing the iris dataset

Plot petal length vs petal width.

library(ggplot2)

# Plot petal length vs width for dataset, distinguish species by color
p <- ggplot(data = iris,
            aes(x = Petal.Width,
                y = Petal.Length,
                color = Species)) +
     geom_point()

# Display plot
p

Chapter 2.4 - iris dataset: petal length vs petal width, species distinguished by color

How does the SVM algorithm deal with multiclass problems?

SVMs are essentially binary classifiers.
Can be applied to multiclass problems using the following voting strategy:
- Partition the data into subsets containing two classes each.
- Solve the binary classification problem for each subset.
- Use majority vote to assign a class to each data point.
Called one-against-one classification strategy.

Building a multiclass linear SVM

Build a linear SVM for the iris dataset
- 80/20 training / test split (seed 10), default cost

library(e1071)

# Build model
svm_model <- svm(Species ~ ., 
                data = trainset, 
                type = "C-classification", 
                kernel = "linear")

Calculate accuracy

pred_train <- predict(svm_model, trainset)
mean(pred_train == trainset$Species)

0.9756098

pred_test <- predict(svm_model, testset)
mean(pred_test == testset$Species)

0.962963

Time to practice!

Support Vector Machines in R