Iris_presentation

Iris Data - ANNetwork

• Debora Drucker,
• Jorge Gomez,
• Henrique Mello,
• Juliano Melis

Objective

• Improve Artificial Neural Network model to classify Iris species

First Step

• It is important to make a F/T data to perform this classification procedure

data(iris)
iris$setosa<-iris$Species=='setosa'
iris$virginica<-iris$Species=='virginica'
iris$versicolor<-iris$Species=='versicolor'
head(iris)

##   Sepal.Length Sepal.Width Petal.Length Petal.Width Species setosa
## 1          5.1         3.5          1.4         0.2  setosa   TRUE
## 2          4.9         3.0          1.4         0.2  setosa   TRUE
## 3          4.7         3.2          1.3         0.2  setosa   TRUE
## 4          4.6         3.1          1.5         0.2  setosa   TRUE
## 5          5.0         3.6          1.4         0.2  setosa   TRUE
## 6          5.4         3.9          1.7         0.4  setosa   TRUE
##   virginica versicolor
## 1     FALSE      FALSE
## 2     FALSE      FALSE
## 3     FALSE      FALSE
## 4     FALSE      FALSE
## 5     FALSE      FALSE
## 6     FALSE      FALSE

Second Step

• Separate a train & a test dataset

id_train <- sample(x = nrow(iris), size = nrow(iris)*0.5)
iris_train <- iris[id_train,]
iris_test <- iris[-id_train,]
table(iris_test$Species)

## 
##     setosa versicolor  virginica 
##         24         25         26

table(iris_train$Species)

## 
##     setosa versicolor  virginica 
##         26         25         24

Third step

• As performed by Raphael in the class

colnames(iris_test)

## [1] "Sepal.Length" "Sepal.Width"  "Petal.Length" "Petal.Width" 
## [5] "Species"      "setosa"       "virginica"    "versicolor"

if(!require(neuralnet)){install.packages('neuralnet')}

## Loading required package: neuralnet

## Warning: package 'neuralnet' was built under R version 3.3.3

iris_net <- neuralnet(setosa + versicolor + virginica ~ 
                        Sepal.Width + Sepal.Length, 
                      data=iris_train, hidden = 3, 
                      err.fct = 'ce', linear.output = F,
                      lifesign = 'minimal', # 'none', 'full' or 'minimal'
                      rep=2, stepmax = 1e+06) # read help

## hidden: 3    thresh: 0.01    rep: 1/2    steps:  192353  error: 18.92344 time: 28.19 secs
## hidden: 3    thresh: 0.01    rep: 2/2    steps:  231344  error: 24.67674 time: 34.16 secs

plot(iris_net, rep='best')

Checking the model

• Performing a confusion matrix

iris_prediction <- compute(iris_net, 
                           iris_test[, c("Sepal.Width","Sepal.Length")])
idx <- apply(iris_prediction$net.result, 1, FUN=which.max)
predicted<-c('setosa','virginica','versicolor')[idx]
table(predicted, iris_test$Species)

##             
## predicted    setosa versicolor virginica
##   setosa         23          1         0
##   versicolor      0         11        16
##   virginica       1         13        10

Verifying the dataset

## Loading required package: ggplot2

Verifying the dataset

Verifying the dataset

Verifying the dataset

Improving the model

• First attempt

iris_net1 <- neuralnet(setosa + versicolor + virginica ~ 
                        Sepal.Width + Sepal.Length+ Petal.Width + Petal.Length, 
                      data=iris_train, hidden = 3, 
                      err.fct = 'ce', linear.output = F,
                      lifesign = 'minimal', # 'none', 'full' or 'minimal'
                      rep=2, stepmax = 1e+06) # read help

## hidden: 3    thresh: 0.01    rep: 1/2    steps:   13671  error: 0.00109  time: 1.99 secs
## hidden: 3    thresh: 0.01    rep: 2/2    steps:  110721  error: 0.00418  time: 15.86 secs

Improving the model

• First attempt

iris_prediction <- compute(iris_net1, 
                           iris_test[, c("Sepal.Width","Sepal.Length","Petal.Width","Petal.Length")])
idx <- apply(iris_prediction$net.result, 1, FUN=which.max)
predicted<-c('setosa','versicolor','virginica')[idx]
table(predicted, iris_test$Species)

##             
## predicted    setosa versicolor virginica
##   setosa         24          0         0
##   versicolor      0         22        12
##   virginica       0          3        14

Improving the model

• First attempt

plot(iris_net1, rep='best')

Improving the model

• Second attempt

iris_net2 <- neuralnet(setosa + versicolor + virginica ~ 
                         Sepal.Width + Sepal.Length+ Petal.Width + Petal.Length, 
                       data=iris_train, hidden = c(2,10), 
                       err.fct = 'ce', linear.output = F,
                       lifesign = 'minimal', # 'none', 'full' or 'minimal'
                       rep=2, stepmax = 1e+06) # read help

## hidden: 2, 10    thresh: 0.01    rep: 1/2    steps:   25582  error: 0.00017  time: 6.33 secs
## hidden: 2, 10    thresh: 0.01    rep: 2/2    steps:    9086  error: 0.00073  time: 2.26 secs

Improving the model

• Second attempt

iris_prediction <- compute(iris_net2, 
                           iris_test[, c("Sepal.Width","Sepal.Length","Petal.Width","Petal.Length")])
idx <- apply(iris_prediction$net.result, 1, FUN=which.max)
predicted<-c('setosa','versicolor','virginica')[idx]
table(predicted, iris_test$Species)

##             
## predicted    setosa versicolor virginica
##   setosa         24          1         0
##   versicolor      0         22        10
##   virginica       0          2        16

Improving the model

• Second attempt

plot(iris_net2, rep='best')

Improving the model

• Giving up. Let’s try LDA!

require(MASS)

## Loading required package: MASS

iris_lda <- lda(Species~Sepal.Width + Sepal.Length+ Petal.Width + Petal.Length,data=iris_train)
iris_lda_prediction<-predict(iris_lda,iris_test)$posterior
id_lda<-apply(iris_lda_prediction, 1, FUN=which.max)
predicted_lda<-c('setosa','versicolor','virginica')[id_lda]

Improving the model

• Let’s try LDA!

table(predicted_lda, iris_test$Species)

##              
## predicted_lda setosa versicolor virginica
##    setosa         24          0         0
##    versicolor      0         25         5
##    virginica       0          0        21

Conclusion

• Old methods might be useful sometimes…