Zillow Gradient Boosted Machine

Data Preprocessing

property = read_csv(file = "properties_2016.csv", header = T, sep = ",")

Error in read_csv(file = "properties_2016.csv", header = T, sep = ",") : 
  unused arguments (header = T, sep = ",")

Change Variable Types

### Some variables are seen as numerical values, but they should be categorical, hence transformation 
change_numerical_to_factor = function(data, variable_names) {
  for (i in 1:length(variable_names)) {
    index = which(names(data) == cate_var[i])
    if (length(index) > 0) {
      data[,index] = as.factor(data[,index])
    }
  }
  return(data)
}
cate_var = c( "buildingqualitytypeid", "fips", "heatingorsystemtypeid", "propertylandusetypeid", 'censustractandblock', 
              'regionidcity', 'regionidcounty', 'regionidcity', "unitcnt")
train = change_numerical_to_factor(train, cate_var)
property = change_numerical_to_factor(property, cate_var)

Feature Engineering

#### month of the year shows predicative power, so, use month column 
#Take in data and create features.  Use this for the train and total properties files
#This is EXCLUDING the month features which need to be added later
make_features = function(data) {
  data %>% 
    mutate(#transactiondate = as.Date(as.character(transactiondate)),
           #month = as.factor(month(transactiondate)),
           N_value_ratio = taxvaluedollarcnt/taxamount,
           N_living_area_prop = calculatedfinishedsquarefeet/lotsizesquarefeet) %>%
    
    group_by(regionidcity) %>%
    
    mutate( N_Avg_structuretaxvalue = mean(structuretaxvaluedollarcnt),
            N_city_count = n()) %>%
    ungroup() %>%
    
    mutate( N_Dev_structuretaxvaluedolarcnt = abs(structuretaxvaluedollarcnt - N_Avg_structuretaxvalue)/N_Avg_structuretaxvalue,
            N_tax_score = taxvaluedollarcnt*taxamount)  %>%
    
    group_by(regionidzip) %>%
    
    mutate(N_zip_count = n()) %>%
    ungroup() %>%
    
    dplyr::select(- rawcensustractandblock,
           - propertyzoningdesc,
           - censustractandblock,
           - assessmentyear) ### Remove transactiondate column 
  
  #data$N_month = as.factor(data$N_month)
}
train = train %>%
  mutate(transactiondate = as.Date(as.character(transactiondate))) %>% 
    mutate(month = as.factor(month(transactiondate)))  %>% 
    select(-one_of("transactiondate"))

Error in mutate_impl(.data, dots) : 
  Evaluation error: object 'transactiondate' not found.

Remove NA values

prop2 = property %>% dplyr::select(-drop.column)   

Error in -drop.column : invalid argument to unary operator

Remove outliers observations

Model

number_trees = 200
gbmModel <- gbm(logerror ~ ., 
                distribution="gaussian", 
                # var.monotone=c(0,0,0,0,0,0),  # -1: monotone decrease,
                # +1: monotone increase,
                #  0: no monotone restrictions
                interaction.depth = 5,          ### 1 means additive model, 2 is two-way interaction 
                # n.minobsinnode = 10,         ### minimum number of observations in the trees terminal nodes.
                n.cores=detectCores()/2,      ### Number of cores for parallasation
                bag.fraction = 0.8,          ### Every time only x% of the sample are selected to make the tree
                n.trees = number_trees,              ### Number of trees 
                shrinkage = 0.033,           ### There is a shrinkage parameter to avoid over-fitting 
                data = train)

Predictions