zillow_rf_model.py

"""
This is one component I have been integrating into my model for predicting the the zestimate error (I'm just outside top 100).
On its own the random forest is not the greatest of predictive models (~0.0648) and it is greatly outperformed by
a standalone XGBoost model or standalone LightGBM model (both found in other kernels already posted). I have been blending multiple 
models together and I have had success with integrating the random forest regression outputs from this script at a low weight.
I hope some other people can integrate this small snippet of code into their solutions and improve their results as well. 


"""


import numpy as np
import pandas as pd
from sklearn.ensemble import RandomForestRegressor
from sklearn.cross_validation import cross_val_score
from sklearn.preprocessing import LabelEncoder


print('reading train and test data')
#id and 57 predictor columns such as number of bedrooms and stufff like that
properties = pd.read_csv('../input/properties_2016.csv')
#train has the logerror, a parcelid and a transaction date
train = pd.read_csv('../input/train_2016_v2.csv')
#read in the test_ids needed for submission
test_ids = pd.read_csv('../input/sample_submission.csv')

#fill median for numeric columns
for c in properties.dtypes[properties.dtypes == 'float64'].index.values:
    properties[c].fillna(properties[c].median(skipna=True), inplace = True)
#fill -1 for categorical columns
for c in properties.columns:
    properties[c]=properties[c].fillna(-1)
    if properties[c].dtype == object:
        lbl = LabelEncoder()
        lbl.fit(list(properties[c].values))
        properties[c] = lbl.transform(list(properties[c].values))

#merge the training data with the properties
train_df = train.merge(properties, how='left', on='parcelid')
#drop the id from the test dataframe
x_test = properties.drop(['parcelid'], axis=1)
#drop id, data and logerror(response) from the df to make the training df
x_train = train_df.drop(['parcelid', 'logerror','transactiondate'], axis=1)
#retrieve just the logerror of the training set
y_train = train_df["logerror"].values.astype(np.float32)


#######
# turn the general pandas dataframes into numpy arrays for use in the random forest regressions
#######
x_train = x_train.values.astype(np.float32, copy=False)
x_test = x_test.values.astype(np.float32, copy=False)
z_rf = RandomForestRegressor(max_depth= 8, n_estimators = 100)

z_rf.fit(x_train, y_train)

rf_y_test = z_rf.predict(x_test)

"""
#uncomment this section and run locally in order to optimize the paramaters, you can add others as well.
param_grid = {
                 'n_estimators': [500, 1000, 1500, 2000],
                 'max_depth': [5, 7, 9]
             }

grid = GridSearchCV(z_rf, param_grid, cv=10)

grid.fit(x_train, y_train)

print(grid.best_score_)
print(grid.best_params_)
"""


test_columns = ['201610','201611','201612','201710','201711','201712']

for i in test_columns:
    test_ids[i] = [float(format(house, '.4f')) for house in rf_y_test]
    
test_ids.to_csv('cam_rf_model_component.csv', index=False)