In brian-j-smith/MachineShop: Machine Learning Models and Tools

IC Home Prices

In this example, a generalized boosted model is developed to predict the prices of homes sold in Iowa City, IA during 2005-2008.

library(MachineShop)

stats <- list(
  list("Number of homes" = ~ length(sale_amount)),
  "sale_amount" = list("Median (Range)" = ~ median_range(sale_amount, "$",
                                                         big.mark = ",",
                                                         big.interval = 3)),
  "sale_year" = list("Median (Range)" = ~ median_range(sale_year)),
  "sale_month" = list("Median (Range)" = ~ median_range(sale_month)),
  "built" = list("Median (Range)" = ~ median_range(built)),
  "style" = list("Home" = ~ n_perc(style == "Home"),
                 "Condo" = ~ n_perc(style == "Condo")),
  "construction" = list(
    "1 1/2 Story Frame" = ~ n_perc(construction == "1 1/2 Story Frame"),
    "1 Story Brick" = ~ n_perc(construction == "1 Story Brick"),
    "1 Story Condo" = ~ n_perc(construction == "1 Story Condo"),
    "1 Story Frame" = ~ n_perc(construction == "1 Story Frame"),
    "2 Story Brick" = ~ n_perc(construction == "2 Story Brick"),
    "2 Story Condo" = ~ n_perc(construction == "2 Story Condo"),
    "2 Story Frame" = ~ n_perc(construction == "2 Story Frame"),
    "Split Foyer Frame" = ~ n_perc(construction == "Split Foyer Frame"),
    "Split Level Frame" = ~ n_perc(construction == "Split Level Frame")),
  "base_size (sq ft)" = list("Median (Range)" = ~ median_range(base_size)),
  "garage1_size (sq ft)" = list("Median (Range)" = ~ median_range(garage1_size)),
  "garage2_size (sq ft)" = list("Median (Range)" = ~ median_range(garage2_size)),
  "lot_size (sq ft)" = list("Median (Range)" = ~ median_range(lot_size)),
  "bedrooms" = list("1-2" = ~ n_perc(bedrooms %in% 1:2),
                    "3" = ~ n_perc(bedrooms == 3),
                    "4" = ~ n_perc(bedrooms == 4),
                    "5+" = ~ n_perc(bedrooms > 4)),
  "basement" = list("Yes" = ~ n_perc(basement == "Yes"),
                    "No" = ~ n_perc(basement == "No")),
  "ac" = list("Yes" = ~ n_perc(ac == "Yes"),
              "No" = ~ n_perc(ac == "No")),
  "attic" = list("Yes" = ~ n_perc(attic == "Yes"),
                 "No" = ~ n_perc(attic == "No")),
  "lon" = list("Median (Range)" = ~ median_range(lon)),
  "lat" = list("Median (Range)" = ~ median_range(lat)))

summary_kbl(stats, ICHomes)

Training Set Analysis

## Analysis libraries
library(MachineShop)
library(ggplot2)

## Training and test sets
set.seed(123)
train_indices <- sample(nrow(ICHomes), nrow(ICHomes) * 2 / 3)
trainset <- ICHomes[train_indices, ]
testset <- ICHomes[-train_indices, ]

## Model formula
fo <- sale_amount ~ .

## Boosted regression model tuned with the training set
model_fit <- TunedModel(GBMModel, grid = 5) %>% fit(fo, data = trainset)

## Variable importance
vi <- varimp(model_fit)
plot(vi)

## Performance plotted over the grid points
(tuned_model <- as.MLModel(model_fit))
plot(tuned_model, type = "line")

Generalization Performance

## Test set observed and predicted sale amounts
obs <- response(model_fit, newdata = testset)
pred <- predict(model_fit, newdata = testset)

## Test set performance
performance(obs, pred)

Calibration Curve

cal <- calibration(obs, pred, breaks = NULL)
plot(cal, se = TRUE)

Partial Dependence Plots

## Marginal predictor effects
pd <- dependence(model_fit, select = c(base_size, built, basement))
plot(pd)

## Spatial distribution
pd <- dependence(model_fit, select = c(lon, lat), interaction = TRUE, n = 25)
df <- cbind(pd$Predictors, sale_amount = pd$Value)
ggplot(df) +
  stat_summary_2d(aes(lon, lat, z = sale_amount), binwidth = 0.006) +
  geom_point(aes(lon, lat), data = trainset) +
  labs(x = "Longitude", y = "Latitude", fill = "Sale Amount")