In vasterlund/Lab7_Albin_Eric: Lab7

library(mlbench)
library(caret)
library(MASS)
library(leaps)
library(Lab7SpaghettiBolognese)
data("BostonHousing")

1)

We will devide the BostonHousing dataset in to a training and a test set.

set.seed(12345)

training <- createDataPartition(BostonHousing$tax,p = 0.7)

train_data <- BostonHousing[training$Resample1, ]
test_data <- BostonHousing[-training$Resample1, ]

2)

This is our full linear model with the respons variable tax.

model_lm <- train(tax ~ .  ,data = train_data, method = "lm")
sol <- summary(model_lm)
sol

A lot of parameters is un-significant. So we need to reduce our model. We do it by forward selection.

train_leapForward(data = train_data, formula = tax ~ .)$model

model_lm <- train(tax ~ zn + indus + rad + medv  ,data = train_data, method = "lm")
summary(model_lm)

We can see that the model choosen contains the expnatory variables: zn, indus, rad and medv.

3)

Now we want to evaluate the model and compare it with the other model in the forward selection.

train_leapForward(data = train_data, formula = tax ~ .)$leapForward

The final model have smalest RMSE, and highest adjusted r-square which makes the model the best one.

4)

In this task we will test different $\lambda$ for our ridereg() function. We had to do a function with our train for our vingnette to actually work.

ridgereg_train(lambda = c(3,4,6,8,5), normalize = TRUE)

We test for the values $\lambda = 3,4,6,8, 5$ and we get that the best lambda is 8 when looking at the RMSE.

5)

We will do a 10-fold cross-validation in the training data set.

ridgereg_train(lambda = c(3,4,6,8,5),cross_val = TRUE,normalize = TRUE)

We perform the cross-validation 10 times for $\lambda = 3,4,6,8, 5$. We set the normalize to FALSE do we can compare the RMSE values.

6)

We can see that by just looking at the RSME value that the model with the lowest is from the linear model.

summary(model_lm)

We recive a print-out from the best model of the three. There would probably be more investigation of what $\lambda$ would be the best to get best model.

vasterlund/Lab7_Albin_Eric documentation built on May 7, 2019, 3:57 a.m.