///

library(tidyverse)
library(tictoc); library(doParallel)
library(stringr); library(lubridate); library(inspectdf); library(PerformanceAnalytics); library(fastDummies)
library(caret); library(mlbench); library(xgboost); library(earth); library(ranger)
library(OpenStreetMap)
library(patchwork); library(hrbrthemes); theme_set(theme_ipsum())

# beijing_housing -> bh
bh <- read_csv("data/new.csv", locale = readr::locale(encoding = "latin1")) %>%
  select(-url, -id, -Cid) # removing ids

# save test and make sample df
set.seed(23)
bh <- bh[sample(nrow(bh)),] # random shuffle rows
df_splits <- bh %>%
  mutate(split_flag = sample(1:2, size = dim(bh)[1], prob=c(0.1, 0.9), replace=TRUE)) %>%
  mutate(split_flag = ifelse(split_flag == 1, "test", "dev")) %>%
  group_by(split_flag) %>%
  group_split()

bh_dev <- df_splits[[1]]
bh_test <- df_splits[[2]]
write_csv(bh_test, "bh_test.csv")
bh <- bh_dev
bh_sample <- bh %>% sample_n(0.02 * dim(bh)[1])

colnames(bh)

 [1] "Lng"                 "Lat"                 "tradeTime"          
 [4] "DOM"                 "followers"           "totalPrice"         
 [7] "price"               "square"              "livingRoom"         
[10] "drawingRoom"         "kitchen"             "bathRoom"           
[13] "floor"               "buildingType"        "constructionTime"   
[16] "renovationCondition" "buildingStructure"   "ladderRatio"        
[19] "elevator"            "fiveYearsProperty"   "subway"             
[22] "district"            "communityAverage"    "split_flag"         

print(paste("Development dataset size: ", dim(bh)))

[1] "Development dataset size:  286904" "Development dataset size:  24"    

knitr::kable(t(head(bh)))

inspect_types(bh, show_plot = TRUE)

The `show_plot = TRUE` is deprecated and will be removed in a future version.  The `show_plot()` function should be used instead.  For more info, check out the help file ?show_plot()

inspect_na(bh)

ggplot(bh, aes(DOM)) +
  geom_histogram(bins = 100) +
  geom_vline(xintercept = median(bh$DOM, na.rm=TRUE), colour = "red")

ggplot(bh, aes(price)) + geom_histogram(bins=100)

PerformanceAnalytics::chart.Correlation(select(bh_sample, DOM, followers, price, square, ladderRatio, totalPrice, communityAverage))

weekly_avg <- bh %>%
  group_by(week = lubridate::floor_date(as.Date(tradeTime), unit = "week")) %>%
  summarize(avg_price = mean(price)) %>%
  ungroup() %>%
  filter(year(week) > 2011)

ggplot(weekly_avg, aes(week, avg_price)) +
  geom_line() +
  geom_smooth() +
  labs(y = "Average weekly price", x = "Week")

map <- openmap(upperLeft = c(max(bh$Lat),min(bh$Lng)),
               lowerRight = c(min(bh$Lat),max(bh$Lng)),
               zoom = NULL,
               type = c("osm"))
map.latlon <- openproj(map, projection = "+proj=longlat +ellps=WGS84 +datum=WGS84 +no_defs")

autoplot(map.latlon) +
  geom_point(data = bh_sample, aes(Lng, Lat, colour = price), alpha = 0.6) + 
  labs(title = "Beijing - house price by location", x = " ", y = " ", colour = "Price") + 
  theme(plot.margin = unit(c(0,0,0,0),"cm"))

p1 <- autoplot(map.latlon) +
  geom_point(data = bh_sample, aes(Lng, Lat, colour = as.factor(district)), alpha = 0.6) + 
  labs(title = "Price distribution by district", x = " ", y = " ", colour = "District") + 
  theme(plot.margin = unit(c(0,0,0,0),"cm"), legend.position = "none")

p2 <- ggplot(bh, aes(as.factor(district), price, fill = as.factor(district))) + 
  geom_boxplot(outlier.shape = NA) + 
  scale_x_discrete(position = "top") +
  labs(title = " ", x = "District", y = "Price", colour = " ") + 
  coord_flip(ylim=c(0, 100000)) +
  theme(legend.position = "none", plot.margin = unit(c(0,0,0,0),"cm"))

p1 + p2 + plot_layout(ncol = 2)

base_fit <- lm(price ~ ., data = select(bh, -totalPrice, -communityAverage, -followers, -DOM, -split_flag) %>% na.omit)
cat(paste0("R-squared: ", round(summary(base_fit)$adj.r.squared, 2),
           "\nRMSE: ", round(sqrt(mean(base_fit$residuals^2)), 2),
           "\nMAE: ", round(mean(abs(base_fit$residuals)), 2)))

R-squared: 0.46
RMSE: 15909.25
MAE: 11758.45

date_features <- function(dates, as_factors = FALSE){
  "Takes a vector of dates and returns a dataframe including year, quarter_of_year, month_of_year, week_of_year, week_of_month, day_of_month, weekday and weekend"
  
  monthweeks <- function(x) {
    ceiling(as.numeric(format(x, "%d"))/7)
  }
  dates <- as.data.frame(dates)
  colnames(dates) <- c("date")
  dates_features <- dates %>% mutate(year = lubridate::year(date)) %>% 
    mutate(quarter_of_year = lubridate::quarter(date)) %>% 
    mutate(month_of_year = lubridate::month(date)) %>%
    mutate(month_name = lubridate::month(date, label = TRUE)) %>%
    mutate(week_of_year = lubridate::week(date)) %>% 
    mutate(week_of_month = monthweeks(date)) %>% mutate(day_of_month = lubridate::day(date)) %>% 
    mutate(day_of_year = lubridate::yday(date)) %>% mutate(weekday = lubridate::wday(date)) %>% 
    mutate(weekday_name = lubridate::wday(date, label = TRUE)) %>% 
    mutate(weekend = ifelse(weekday %in% c(1, 7), 1, 0)) %>% 
    mutate(ymd_format = lubridate::ymd(paste(year, month_of_year, day_of_month, sep = "-")))
  if (as_factors) {
    dates_features[colnames(dates_features)[-1]] <- lapply(dates_features[colnames(dates_features)[-1]], 
      factor)
  }
  return(dates_features)
}

tradetime_features <- date_features(bh$tradeTime, as_factors = FALSE) %>% select(-date, -ymd_format, -weekday, -month_of_year, -day_of_year)

t(head(tradetime_features))

                1      2      3      4      5      6     
year            "2016" "2017" "2015" "2013" "2015" "2013"
quarter_of_year "1"    "3"    "1"    "3"    "4"    "1"   
month_name      "lut"  "lip"  "mar"  "wrz"  "lis"  "sty" 
week_of_year    " 8"   "30"   "12"   "37"   "45"   " 5"  
week_of_month   "3"    "4"    "3"    "2"    "1"    "5"   
day_of_month    "21"   "23"   "21"   "13"   " 5"   "30"  
weekday_name    "nie"  "nie"  "sob"  "pią"  "czw"  "śro" 
weekend         "1"    "1"    "1"    "0"    "0"    "0"   

bh_processed <- bh %>%
  rename(floor_character = `floor`) %>%
  cbind(tradetime_features) %>%
  mutate(floor_number = str_extract(floor_character, "(\\d+)(?: \\d+)*$")) %>%
  filter(as.numeric(year) > 2011) %>%
  map_at(c("livingRoom", "drawingRoom", "kitchen", "bathRoom", "buildingType" , "buildingStructure", "renovationCondition", "fiveYearsProperty", "elevator", "subway", "district", "quarter_of_year", "weekend", "weekday_name", "month_name", "year"), as.factor) %>%
  map_at(c("Lng", "Lat", "DOM", "followers", "price", "square", "constructionTime", "ladderRatio", "floor_number", "constructionTime", "week_of_year", "day_of_month"), as.numeric) %>%
  as_tibble() %>%
  select(-floor_character, -tradeTime, -totalPrice, -communityAverage, -DOM, -followers, -split_flag) %>%
  filter(buildingType %in% c(1, 2, 3, 4)) %>%
  filter(ladderRatio < quantile(bh$ladderRatio, 0.99)) %>%
  filter(price > 10000) %>%
  filter(constructionTime > 1800) %>%
  mutate(naConstruction = as.factor(ifelse(is.na(constructionTime), "na_ct", "ok_ct"))) %>%
  mutate(constructionTime = replace_na(constructionTime, 0)) %>%
  na.omit()

stopifnot(map(bh_processed, class)[1] %in% c("factor", "ordered", "numeric"))
stopifnot(sum(is.na(bh_processed)) == 0)

# faster then caret defaults
bh_factors <- select_if(bh_processed, is.factor)
dummies <- select_if(fastDummies::dummy_cols(bh_factors), is.integer)
bh_processed <- select_if(bh_processed, is.numeric) %>% cbind(dummies)
bh_processed <- cbind(bh_processed[,1:10], map_df(bh_processed[11:dim(bh_processed)[2]], as.factor))

set.seed(23)
tv_split <- createDataPartition(bh_processed$price, p = 0.8, list=FALSE)
bh_train <- bh_processed[tv_split,]
bh_val <- bh_processed[-tv_split,]

x_train <- bh_train %>% select(-price)
y_train <- bh_train %>% select(price)
x_val <- bh_val %>% select(-price)
y_val <- bh_val %>% select(price)

# caret preprocessing would automatically scale dummies if we keep them as numeric and skip other operations if we keep them as factors -> two preproc steps

tic("caret preproc")
preproc_numeric <- preProcess(x_train, method = c("center", "scale", "YeoJohnson", "knnImpute"))
x_train <- predict(preproc_numeric, x_train)
x_train <- map_df(x_train, as.numeric)
preproc_all <- preProcess(x_train, method = c("nzv", "corr", "conditionalX"))
x_train <- predict(preproc_all, x_train)

x_val <- predict(preproc_numeric, x_val)
x_val <- map_df(x_val, as.numeric)
x_val <- predict(preproc_all, x_val)

write_rds(x_train, "data/x_train.rds")
write_rds(x_val, "data/x_val.rds")

toc()

caret preproc: 21.237 sec elapsed

print(preproc_all)

Created from 205238 samples and 83 variables

Pre-processing:
  - conditionalX (42)
  - ignored (0)
  - removed (41)

print(preproc_numeric)

Created from 205238 samples and 83 variables

Pre-processing:
  - centered (9)
  - ignored (74)
  - 5 nearest neighbor imputation (9)
  - scaled (9)
  - Yeo-Johnson transformation (6)

Lambda estimates for Yeo-Johnson transformation:
-0.19, 0.77, 0.72, 0.66, 0.8, -0.26

knitr::kable(t(head(x_train)))

dim(x_train)

[1] 205238     42

gc()

            used   (Mb) gc trigger   (Mb)  max used   (Mb)
Ncells   3138924  167.7    4703850  251.3   4703850  251.3
Vcells 153017995 1167.5  316973979 2418.4 316915827 2417.9

tic("Training models")
set.seed(23)
training_data <- cbind(y_train, x_train)
methods <- c("lasso", "earth", "brnn", "ranger")
fitControl <- trainControl(method = "repeatedcv",
                           number = 5,
                           repeats = 2,
                           search = "grid",
                           allowParallel = TRUE,
                           verboseIter = TRUE)

#cl <- makePSOCKcluster(2)
#registerDoParallel(cl)
models_metrics <- list()
for (m in methods) {

  model <- caret::train(price ~ .,
                        data = training_data,
                        method = m,
                        trControl = fitControl,
                        importance = 'impurity',
                        na.action = na.pass)

  saveRDS(model, paste0("models/", m, ".rds"))
  models_metrics[[m]] <- model$resample
  gc()
}

#stopCluster(cl)
toc()

models <- list.files("models/")
models_metrics <- list()
for (model in models) {
  fit <- read_rds(paste0("models/", model))
  models_metrics[[fit$method]] <- fit$resample
}

df_metrics <- do.call(what = rbind, args = models_metrics) %>%
  rownames_to_column() %>%
  separate(rowname, into = c("model_name", "iteration")) %>%
  filter(model_name %in% c("lasso", "earth", "brnn", "ranger")) %>%
  mutate(model_name = ifelse(model_name == "ranger", "random_forest", model_name))

p1 <- ggplot(df_metrics, aes(forcats::fct_reorder(model_name, MAE, .desc = TRUE), MAE)) +
  geom_boxplot() + 
  geom_point() +
  labs(title = "Resampled metrics", subtitle = "Mean absolute error", x = " ", y = "MAE") + 
  theme(legend.position = "none", plot.margin = unit(c(0,0,0,0),"cm"))

p2 <- ggplot(df_metrics, aes(forcats::fct_reorder(model_name, Rsquared, .desc = TRUE), Rsquared)) +
  geom_boxplot() + 
  geom_point() +
  labs(subtitle = "Rsquared", x = " ", y = "Rsquared") + 
  theme(legend.position = "none", plot.margin = unit(c(0,0,0,1),"cm"))

p1 + p2 + patchwork::plot_layout(ncol = 2)

rf <- read_rds("models/ranger.rds")
rf$finalModel

Ranger result

Call:
 ranger::ranger(dependent.variable.name = ".outcome", data = x,      mtry = min(param$mtry, ncol(x)), min.node.size = param$min.node.size,      splitrule = as.character(param$splitrule), write.forest = TRUE,      probability = classProbs, ...) 

Type:                             Regression 
Number of trees:                  500 
Sample size:                      205238 
Number of independent variables:  42 
Mtry:                             42 
Target node size:                 5 
Variable importance mode:         impurity 
Splitrule:                        variance 
OOB prediction error (MSE):       24989313 
R squared (OOB):                  0.9464689 

model <- rf
y_hat <- predict(model, newdata = x_val)

preds_df <- as.data.frame(y_hat) %>%
  mutate(y = y_val$price) %>%
  mutate(residuals = y_hat - y,
         sd_residuals = scale(residuals),
         abs_error = abs(residuals))

p1 <- ggplot() + 
  geom_density(aes(y), data = preds_df, colour = "darkgreen", fill = "darkgreen", alpha = 0.5) + 
  geom_density(aes(y_hat), data = preds_df, colour = "darkblue", fill = "darkblue", alpha = 0.5) + 
  labs(x = "price", title="Best model - validation", subtitle="Actuals & predictions distributions") +
  scale_x_comma()

p2 <- ggplot(preds_df, aes(y, y_hat)) +
  geom_point(alpha = 0.5) +
  geom_smooth() + 
  scale_x_comma() +
  scale_y_comma() +
  labs(x = "Price", y="Predicted price", subtitle="Actuals vs predictions")

p3 <- ggplot(preds_df, aes(abs_error)) +
  geom_histogram(bins = 100) +
  scale_x_comma(limits = c(0, 40000)) +
  labs(x = "Absolute error", subtitle = "Absolute error distribution")

p4 <- ggplot(preds_df, aes(y_hat, sd_residuals)) + 
  geom_point(alpha = 0.5) + 
  geom_hline(yintercept = 0, colour="red", linetype = "dashed") + 
  labs(x = "Predicted price", y = "Standarized residuals", subtitle = "Residuals")

p1 + p2 + p3 + p4 + plot_layout(nrow = 2, ncol = 2)

cat(paste0("Median price in the validation set: ", median(y_val$price),
          "\nMean absolute error: ", round(mean(model$resample$MAE), 2)))

Median price in the validation set: 39330.5
Mean absolute error: 3199.81

features_importance <- varImp(model)$importance %>%
  rownames_to_column() %>%
  select(feature = rowname, gini_importance = Overall) %>%
  arrange(desc(gini_importance)) %>%
  head(15)

ggplot(features_importance, aes(forcats::fct_reorder(feature, gini_importance), gini_importance)) + 
  geom_point(size = 2) + 
  coord_flip() +
  labs(x = " ", y = "Gini importance (Mean Decrease in Impurity)", subtitle = "Top features used by random forest for splits")

library(lime)

Attaching package: ‘lime’

The following object is masked from ‘package:dplyr’:

    explain

set.seed(23)
selected_apartments <- sample_n(x_val, 6)
explainer <- lime(selected_apartments, model = model)
explanation <- explain(selected_apartments, explainer, n_features = 7)

plot_features(explanation)

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