Download template here
We will just be using the tidymodels today. If the system prompts you to install a package or gives you a “package not found” error, simply run install.packages("packagename") once to install it.
We will use the ames data set from the modeldata library. It can be loaded using the following code
data("ames", package = "modeldata")
ames
we will set up the splits right away.
set.seed(1234)
ames_split <- initial_split(ames)
ames_train <- training(ames_split)
ames_test <- testing(ames_split)
We are still trying to predict the sale price but we will be trying some new techniques. If we plot the location of the houses by their Longitude and Latitude we can get an idea of the neighborhoods
ggplot(ames_train, aes(Longitude, Latitude)) +
geom_point(alpha = 0.5)
We are actually able to visualize the neighbors directly using Neighborhood.
ggplot(ames_train, aes(Longitude, Latitude, color = Neighborhood)) +
geom_point(alpha = 0.5) +
guides(color = "none")
We can also plot the Sale_Price and we see some localized trends. Note that this doesn’t take into account anything related to the size and features of the houses.
ggplot(ames_train, aes(Longitude, Latitude, color = Sale_Price)) +
geom_point(alpha = 0.5) +
scale_color_viridis_c()
We can also look at the date 1-dimensionally, and we see some non-linear effects happening here.
ggplot(ames_train, aes(Longitude, Sale_Price)) +
geom_point()
We have been looking at a lot of different methods this week. Many of these things are available in {recipes} steps. We will explore some of those in this lab. Let us start with step_poly(), this will create a polynomial expansion of the variables.
rec_poly <- recipe(Sale_Price ~ Longitude + Latitude, data = ames_train) %>%
step_poly(Longitude, Latitude)
We will then combine it with a linear regression specification, into a workflow.
lm_spec <- linear_reg()
poly_wf <- workflow(rec_poly, lm_spec)
and we fit the model
poly_wf_fit <- fit(poly_wf, data = ames_train)
We have seen before how we can calculate metrics and other things to validate the model. This time let us do a more visual inspection of the performance. Let us plot the predicted values on the map we created earlier.
augment(poly_wf_fit, new_data = ames_train) %>%
ggplot(aes(Longitude, Latitude, color = .pred)) +
geom_point(alpha = 0.5) +
scale_color_viridis_c()
While this is cool, it can make it hard to see where we are doing well and where we aren’t. So we can plot the residuals. By using a diverging color palette can we see where we are doing well and where we aren’t.
augment(poly_wf_fit, new_data = ames_train) %>%
ggplot(aes(Longitude, Latitude, color = Sale_Price - .pred)) +
geom_point(alpha = 0.5) +
scale_color_gradient2()
Try different non-linear transformations. step_bs() to fit a splines, or step_discretize() or step_cut() to fit step functions. All of these have a hyperparameter you could tune. Try tuning one to see if that improves your model over the default settings.