#!/usr/bin/env python
# coding: utf-8

# Below, I implement a quick and dirty first ML pipeline: I import a subset of the training data, split it into training/test, choose some features and run a random forest on it.

# In[1]:


get_ipython().system(' ls')


# Import subset of training data for 1st model pipeline:

# In[72]:


#Import libraries:
import numpy as np
import pandas as pd
#import xgboost as xgb
import time
train = pd.read_csv("train.csv", nrows = 10000)
test= pd.read_csv("test.csv", nrows = 1000)


# In[73]:


train.head()


# In[74]:


test.head()


# ___
# Plot histograms of numerical features:

# In[75]:


get_ipython().run_line_magic('matplotlib', 'inline')
import matplotlib.pyplot as plt
plt.style.use('ggplot')
pd.DataFrame.hist(train, figsize = [15,15]);


# ___
# Drop non-numerical feautures in order to run a random forest:

# In[76]:


drop_list = []
for (train_name, train_series) in train.iteritems():
    print(train_name, train_series.dtype)
    if train_series.dtype == 'object':
        drop_list.append(train_name)
drop_list.append('hotel_cluster')
print(drop_list)


# ___
# Split data into predictor variables (X) and target variable (Y):

# In[77]:


X = train.drop(drop_list, axis = 1).fillna(-1)
y = train['hotel_cluster']


# Split into test & train sets, run random forest on train set:

# In[78]:


from sklearn.cross_validation import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)
#See here: http://scikit-learn.org/stable/auto_examples/ensemble/plot_forest_importances.html

from sklearn.ensemble import ExtraTreesClassifier, RandomForestClassifier
# Build a forest and compute the feature importances
forest = RandomForestClassifier(n_estimators=10)
forest.fit(X_train, y_train)
score = forest.score(X_test, y_test)


# Print precision:

# In[79]:


from sklearn.metrics import precision_score
precision_score(y_test, forest.predict(X_test), average="micro")


# Note: this is NOT the quite the metric of model performance we need to use for the competition but it's close. I tihnk that they're equal when we make only one prediction for each row. **Discuss**. This score would get us in the top ~50% of current leaderboard.