Input transforms: the "hockey stick" (2)
In the last exercise, you saw that a quadratic model seems to fit the houseprice
data better than a linear model.
In this exercise, you will confirm whether the quadratic model would perform better on out-of-sample data.
Since this dataset is small, you will use cross-validation. The quadratic formula fmla_sqr
that you created in the last exercise and the houseprice
data frame are available for you to use.
For comparison, the sample code will calculate cross-validation predictions from a linear model price ~ size
.
This exercise is part of the course
Supervised Learning in R: Regression
Exercise instructions
- Use
kWayCrossValidation()
to create a splitting plan for a 3-fold cross validation.- You can set the 3rd and 4th arguments of the function to
NULL
.
- You can set the 3rd and 4th arguments of the function to
- Examine and run the sample code to get the 3-fold cross-validation predictions of the model
price ~ size
and add them to the columnpred_lin
. - Get the cross-validation predictions for price as a function of squared size. Assign them to the column
pred_sqr
.- The sample code gives you the procedure.
- You can use the splitting plan you already created.
- Fill in the blanks to pivot the predictions and calculate the residuals.
- Fill in the blanks to compare the RMSE for the two models. Which one fits better?
Hands-on interactive exercise
Have a go at this exercise by completing this sample code.
# houseprice is available
summary(houseprice)
# fmla_sqr is available
fmla_sqr
# Create a splitting plan for 3-fold cross validation
set.seed(34245) # set the seed for reproducibility
splitPlan <- ___
# Sample code: get cross-val predictions for price ~ size
houseprice$pred_lin <- 0 # initialize the prediction vector
for(i in 1:3) {
split <- splitPlan[[i]]
model_lin <- lm(price ~ size, data = houseprice[split$train,])
houseprice$pred_lin[split$app] <- predict(model_lin, newdata = houseprice[split$app,])
}
# Get cross-val predictions for price as a function of size^2 (use fmla_sqr)
houseprice$pred_sqr <- 0 # initialize the prediction vector
for(i in 1:3) {
split <- ___
model_sqr <- lm(___, data = houseprice[split$train, ])
houseprice$___[split$app] <- predict(___, newdata = houseprice[split$app, ])
}
# Pivot the predictions and calculate the residuals
houseprice_long <- houseprice %>%
pivot_longer(cols = c('pred_lin', 'pred_sqr'), names_to = 'modeltype', values_to = 'pred') %>%
mutate(residuals = ___)
# Compare the cross-validated RMSE for the two models
houseprice_long %>%
group_by(modeltype) %>% # group by modeltype
summarize(rmse = ___)