House Price Analysis in R with XGBoosting

library(ggplot2) # Data visualization
library(caret) # CSV file I/O, e.g. the read_csv function
library(dplyr)
library(mice)
library(Hmisc)
library(lubridate)
library(data.table)
library(xgboost)
library(gbm)

# import data
properties <- fread('../input/properties_2016.csv', stringsAsFactors = FALSE)
train <- fread('../input/train_2016_v2.csv',stringsAsFactors = FALSE)

# check NAs
colSums(is.na(properties)) %>% sort(decreasing = T) %>% barchart(pt.cex = 2.5, main = 'Amount of Missing Value ')

# basic data transformation
properties$hashottuborspa <- ifelse(properties$hashottuborspa == 'true', 1, 0)
properties$fireplaceflag <- ifelse(properties$fireplaceflag == 'true', 1, 0)
properties$taxdelinquencyflag <- ifelse(properties$taxdelinquencyflag == 'Y', 1, 0)
properties$propertycountylandusecode <- as.numeric(as.factor(properties$propertycountylandusecode))
properties$propertyzoningdesc <- as.numeric(as.factor(properties$propertyzoningdesc))

# basic transformation : numeric/ integer to factor
properties$buildingqualitytypeid <- as.factor(properties$buildingqualitytypeid)
properties$fips <- as.factor(properties$fips)
properties$heatingorsystemtypeid <- as.factor(properties$heatingorsystemtypeid)
properties$propertylandusetypeid <- as.factor(properties$propertylandusetypeid)
properties$regionidcity <- as.factor(properties$regionidcity)
properties$regionidcounty <- as.factor(properties$regionidcounty)
properties$regionidcity <- as.factor(properties$regionidcity)
properties$unitcnt <- as.factor(properties$unitcnt)

# training set
df <- left_join(train,properties,on = 'parcelid')

# outliers
df <- df %>% filter(logerror <= 0.4 & logerror >= -0.39)

# add new variables
df$age <- as.numeric(2016 - df$yearbuilt)
df$transactiondate <- as.Date(df$transactiondate)
df$month <- as.factor(month(df$transactiondate))
df$weekdays <- as.factor(weekdays(df$transactiondate))
df$transactiondate <- as.numeric(df$transactiondate)

## transformation all to integer/numeric
df1 <- df %>% mutate_if(is.factor,as.integer)

## user-defined mae function
maeSummary <- function (data,
                        lev = NULL,
                        model = NULL) {
  out <- ModelMetrics::mae(data$obs, data$pred)  
  names(out) <- "MAE"
  out
}

## parameter
xgb_grid <- expand.grid(nrounds = c(160),
                        eta = c(0.1),
                        max_depth = c(5),
                        gamma = c(1),
                        colsample_bytree = c(0.8),
                        min_child_weight = c(0.8),
                        subsample = c(0.9))
fitControl <- trainControl(method = "repeatedcv",
                           number = 10,
                           repeats = 3,
                           summaryFunction = maeSummary)

# train data set
train_x <- as.matrix(df1[,-c(1,2,3)])
train_y <- df1[,"logerror"]

# Model
set.seed(1)
xgbtree_mod <- train(x = train_x,
                     y = train_y,
                     method = 'xgbTree',
                     trControl = fitControl,
                     tuneGrid = xgb_grid,
                     metric = 'MAE')
print(xgbtree_mod)   # MAE 0.05256457

properties1 <-data.matrix(properties)
submission <- properties %>%
  mutate("201610"=predict(object=xgbtree_mod, newdata=properties1),
         month=factor("11", levels = levels(train$month)),
         "201611"=predict(object=xgbtree_mod, newdata=properties1),
         month=factor("12", levels = levels(train$month)),
         "201612"=predict(object=xgbtree_mod, newdata=properties1),
         #month=factor("10", levels = levels(train$month)),
         "201710"=0,
         #month=factor("11", levels = levels(train$month)),
         "201711"=0,
         #month=factor("12", levels = levels(train$month)),
         "201712"=0) %>%
  select(parcelid, `201610`, `201611`, `201612`, `201710`, `201711`, `201712`)


options(scipen = 999) ## DO not use scientific notation
write.csv(submission, "submission_test.csv", row.names = FALSE)