In [1]:
from __future__ import division
from IPython.display import display
from matplotlib import pyplot as plt
%matplotlib inline
import numpy as np
import pandas as pd
import random, sys, os, re
import xgboost as xgb
from sklearn.cross_validation import StratifiedKFold
from sklearn.grid_search import RandomizedSearchCV, GridSearchCV
from sklearn.cross_validation import cross_val_predict, permutation_test_score
In [2]:
SEED = 97
scale = False
minmax = False
norm = False
nointercept = False
N_CLASSES = 2
submission_filename = "../submissions/submission_XGBClassifier.csv"
In [3]:
from load_blood_data import load_blood_data
y_train, X_train = load_blood_data(train=True, SEED = SEED,
scale = scale,
minmax = minmax,
norm = norm,
nointercept = nointercept)
In [4]:
StatifiedCV = StratifiedKFold(y = y_train,
n_folds = 10,
shuffle = True,
random_state = SEED)
In [5]:
%%time
SEED = 0
random.seed(SEED)
clf = xgb.XGBClassifier(learning_rate = 0.1,
n_estimators = 1000,
max_depth = 3,
subsample = 1,
colsample_bytree = 1,
gamma = 0,
nthread = 1,
objective = 'binary:logistic',
min_child_weight = 1,
max_delta_step = 0,
base_score = 0.5,
seed = SEED,
silent = True,
missing = None)
# param_grid = dict(n_estimators = [5, 10, 50, 75, 100],
# max_depth = [1, 2, 3, 4, 5, 6, 7],
# subsample = [1, 0.75, 0.5, 0.25],
# colsample_bytree = [1, 0.75, 0.5, 0.25])
# grid_clf = GridSearchCV(estimator = clf,
# param_grid = param_grid,
# n_jobs = 1,
# cv = StatifiedCV,
# # fit_params = dict(eval_metric = 'auc',#logloss', "error", "auc"
# # eval_set = [(X_train, y_train), (X_valid, y_valid)],
# # early_stopping_rounds = 10,
# # verbose = False),
# verbose = 0
# )
# grid_clf.fit(X = X_train, y = y_train)
# print("clf_params = {}".format(grid_clf.best_params_))
# print("score: {}".format(round(grid_clf.best_score_, 4)))
# print
# clf = grid_clf.best_estimator_
clf_params = {'n_estimators': 50, 'subsample': 0.25, 'colsample_bytree': 0.75, 'max_depth': 5}
clf.set_params(**clf_params)
clf.fit(X_train, y_train)
In [6]:
# from sklearn_utilities import GridSearchHeatmap
# GridSearchHeatmap(grid_clf, y_key='learning_rate', x_key='n_estimators')
# from sklearn_utilities import plot_validation_curves
# plot_validation_curves(grid_clf, param_grid, X_train, y_train, ylim = (0.0, 1.05))
In [7]:
%%time
from sklearn_utilities import plot_learning_curve
plot_learning_curve(estimator = clf,
title = None,
X = X_train,
y = y_train,
ylim = (0.0, 1.10),
cv = StratifiedKFold(y = y_train,
n_folds = 10,
shuffle = True,
random_state = SEED),
train_sizes = np.linspace(.1, 1.0, 5),
n_jobs = -1)
plt.show()
In [8]:
%%time
train_preds = cross_val_predict(estimator = clf,
X = X_train,
y = y_train,
cv = StatifiedCV,
n_jobs = -1,
verbose = 0,
fit_params = None,
pre_dispatch = '2*n_jobs')
y_true, y_pred = y_train, train_preds
In [9]:
from sklearn.metrics import confusion_matrix
cm = confusion_matrix(y_true, y_pred, labels=None)
print cm
from sklearn_utilities import plot_confusion_matrix
plot_confusion_matrix(cm, ['Did not Donate','Donated'])
accuracy = round(np.trace(cm)/float(np.sum(cm)),4)
misclass = 1 - accuracy
print("Accuracy {}, mis-class rate {}".format(accuracy,misclass))
In [10]:
from sklearn.metrics import roc_curve
from sklearn.metrics import roc_auc_score
from sklearn.metrics import log_loss
from sklearn.metrics import f1_score
fpr, tpr, thresholds = roc_curve(y_true, y_pred, pos_label=None)
plt.figure(figsize=(10,6))
plt.plot([0, 1], [0, 1], 'k--')
plt.plot(fpr, tpr)
AUC = roc_auc_score(y_true, y_pred, average='macro')
plt.text(x=0.6,y=0.4,s="AUC {:.4f}"\
.format(AUC),
fontsize=16)
plt.text(x=0.6,y=0.3,s="accuracy {:.2f}%"\
.format(accuracy*100),
fontsize=16)
logloss = log_loss(y_true, y_pred)
plt.text(x=0.6,y=0.2,s="LogLoss {:.4f}"\
.format(logloss),
fontsize=16)
f1 = f1_score(y_true, y_pred)
plt.text(x=0.6,y=0.1,s="f1 {:.4f}"\
.format(f1),
fontsize=16)
plt.xlabel('False positive rate')
plt.ylabel('True positive rate')
plt.title('ROC curve')
plt.show()
In [11]:
%%time
score, permutation_scores, pvalue = permutation_test_score(estimator = clf,
X = X_train.values.astype(np.float32),
y = y_train,
cv = StatifiedCV,
labels = None,
random_state = SEED,
verbose = 0,
n_permutations = 100,
scoring = None,
n_jobs = -1)
plt.figure(figsize=(20,8))
plt.hist(permutation_scores, 20, label='Permutation scores')
ylim = plt.ylim()
plt.plot(2 * [score], ylim, '--g', linewidth=3,
label='Classification Score (pvalue {:.4f})'.format(pvalue))
plt.plot(2 * [1. / N_CLASSES], ylim, 'r', linewidth=7, label='Luck')
plt.ylim(ylim)
plt.legend(loc='center',fontsize=16)
plt.xlabel('Score')
plt.show()
# find mean and stdev of the scores
from scipy.stats import norm
mu, std = norm.fit(permutation_scores)
In [12]:
# format for scores.csv file
import re
algo = re.search(r"submission_(.*?)\.csv", submission_filename).group(1)
print("{: <26} , , {:.4f} , {:.4f} , {:.4f} , {:.4f} , {:.4f} , {:.4f}"\
.format(algo,accuracy,logloss,AUC,f1,mu,std))
In [23]:
clf.set_params(**clf_params)
clf.fit(X_train, y_train)
Out[23]:
In [24]:
from load_blood_data import load_blood_data
X_test, IDs = load_blood_data(train=False, SEED = SEED,
scale = scale,
minmax = minmax,
norm = norm,
nointercept = nointercept)
In [25]:
y_pred = clf.predict(X_test)
print(y_pred[:10])
try:
y_pred_probs = clf.predict_proba(X_test)
print(y_pred_probs[:10])
donate_probs = [prob[1] for prob in y_pred_probs]
except Exception,e:
print(e)
donate_probs = [0.65 if x>0 else 1-0.65 for x in y_pred]
print(donate_probs[:10])
In [13]:
assert len(IDs)==len(donate_probs)
f = open(submission_filename, "w")
f.write(",Made Donation in March 2007\n")
for ID, prob in zip(IDs, donate_probs):
f.write("{},{}\n".format(ID,prob))
f.close()
In [ ]: