def xgboost_cv():
train_start_date = '2016-03-05'
train_end_date = '2016-04-06'
test_start_date = '2016-04-11'
test_end_date = '2016-04-16'
sub_start_date = '2016-02-05'
sub_end_date = '2016-03-05'
sub_test_start_date = '2016-03-05'
sub_test_end_date = '2016-03-10'
user_index, training_data, label = make_train_set(train_start_date,
train_end_date,
test_start_date,
test_end_date)
X_train, X_test, y_train, y_test = train_test_split(training_data,
label,
test_size=0.2,
random_state=0)
dtrain = xgb.DMatrix(X_train, label=y_train)
dtest = xgb.DMatrix(X_test, label=y_test)
param = {
'max_depth': 10,
'eta': 0.05,
'silent': 1,
'objective': 'binary:logistic'
}
num_round = 4000
param['nthread'] = 4
param['eval_metric'] = "auc"
plst = param.items()
plst += [('eval_metric', 'logloss')]
evallist = [(dtest, 'eval'), (dtrain, 'train')]
bst = xgb.train(plst, dtrain, num_round, evallist)
sub_user_index, sub_trainning_date, sub_label = make_train_set(
sub_start_date, sub_end_date, sub_test_start_date, sub_test_end_date)
test = xgb.DMatrix(sub_trainning_date)
#y = bst.predict(test)
pred = sub_user_index.copy()
y_true = sub_user_index.copy()
pred['label'] = y
y_true['label'] = label
report(pred, y_true)
def xgboost_cv():
train_start_date = '2016-03-06'
train_end_date = '2016-04-06'
test_start_date = '2016-04-06'
test_end_date = '2016-04-11'
sub_start_date = '2016-03-11'
sub_end_date = '2016-04-11'
sub_test_start_date = '2016-04-11'
sub_test_end_date = '2016-04-16'
user_index, training_data, label = make_train_set(train_start_date, train_end_date, test_start_date, test_end_date)
X_train, X_test, y_train, y_test = train_test_split(training_data, label, test_size=0.2, random_state=0)
del user_index, training_data, label
dtrain=xgb.DMatrix(X_train.values, label=y_train) #todo: missing=-999.0
dtest=xgb.DMatrix(X_test.values, label=y_test)
del X_train, X_test, y_train, y_test
# param = {'max_depth': 10, 'eta': 0.05, 'silent': 1, 'objective': 'binary:logistic'}
param = {'learning_rate' : 0.1, 'n_estimators': 1000, 'max_depth': 3,
'min_child_weight': 5, 'gamma': 0, 'subsample': 1.0, 'colsample_bytree': 0.8,
'scale_pos_weight': 1, 'eta': 0.05, 'silent': 1, 'objective': 'binary:logistic'}
num_round = 4000
param['nthread'] = 6
param['eval_metric'] = "auc"
plst = param.items()
plst += [('eval_metric', 'logloss')]
evallist = [(dtest, 'eval'), (dtrain, 'train')]
bst=xgb.train( plst, dtrain, num_round, evallist)
del dtrain
print('saving model...')
# flag = 'basic'
# bst.save_model('./cache/' + flag + '_model')
sub_user_index, sub_training_data, sub_label = make_train_set(sub_start_date, sub_end_date,
sub_test_start_date, sub_test_end_date, test=True)
test = xgb.DMatrix(sub_training_data.values)
y = bst.predict(test)
pred = sub_user_index.copy()
pred['label'] = y
pred = pred.sort_values(by=['label'], ascending=False).groupby(['user_id'], as_index=False).first()
limits = np.linspace(0, 0.5, 100)
scores = np.zeros((1,100))
count = 0
for i in limits:
print('--------------------------------------------------------------------')
print('limit=%s' % str(i))
p = pred[pred['label'] > i]
scores[:, count] = report(p, sub_label)
count += 1
print('max score : %s\nmax limit : %s' % (np.max(scores), limits[np.argmax(scores)]))
def xgboost_cv2():
train_start_date = '2016-03-05'
train_end_date = '2016-04-06'
test_start_date = '2016-04-06'
test_end_date = '2016-04-11'
sub_start_date = '2016-03-10'
sub_end_date = '2016-04-11'
sub_test_start_date = '2016-04-11'
sub_test_end_date = '2016-04-16'
user_index, training_data, label = make_train_set(train_start_date,
train_end_date,
test_start_date,
test_end_date)
X_train, X_test, y_train, y_test = train_test_split(training_data,
label,
test_size=0.2,
random_state=0)
dtrain = xgb.DMatrix(X_train.values, label=y_train)
dtest = xgb.DMatrix(X_test.values, label=y_test)
param = {
'learning_rate': 0.05,
'n_estimators': 1000,
'max_depth': 10,
'min_child_weight': 1,
'gamma': 0,
'subsample': 0.8,
'colsample_bytree': 0.8,
'scale_pos_weight': 1,
'eta': 0.05,
'silent': 1,
'objective': 'binary:logistic',
'eval_metric': 'auc'
}
num_round = 300
param['nthread'] = 5
# param['eval_metric'] = "auc"
# plst = param.items()
# plst += [('eval_metric', 'logloss')]
evallist = [(dtest, 'eval'), (dtrain, 'train')]
bst = xgb.train(param, dtrain, num_round, evallist)
sub_user_index, sub_trainning_date, sub_label = make_train_set(
sub_start_date, sub_end_date, sub_test_start_date,
sub_test_end_date) # use this data to see the offline score
test = xgb.DMatrix(sub_trainning_date.values)
y = bst.predict(test)
pred = sub_user_index.copy()
y_true = get_labels_8(
sub_test_start_date,
sub_test_end_date) # during the test date, real label for cate 8
# y_true = sub_user_index.copy()
pred[
'label'] = y # add the new column which is the predict label for the test date
# print(pred[(pred.label >= 0.12)].shape)
# print("y_true:")
# print(y_true)
# pred = pred[(pred.label >= 0.35)]
# print(len(pred))
# print(pred)
ans = []
for i in range(0, 30):
pred = sub_user_index.copy()
pred['label'] = y
pred = pred[pred.label >= i / 100]
# print(pred)
rep = report(pred, y_true)
print('%s : score:%s' % (i / 100, rep))
ans.append([i / 100, rep])
print('ans:%s' % ans)
threshold = sorted(ans, key=getKey, reverse=True)[0][0]
bestscore = sorted(ans, key=getKey, reverse=True)[0][1]
print('best threshold:%s' % threshold)
print('best score:%s' % bestscore)
def gbdt_cv():
train_start_date = '2016-03-05'
train_end_date = '2016-04-06'
test_start_date = '2016-04-06'
test_end_date = '2016-04-11'
sub_start_date = '2016-03-10'
sub_end_date = '2016-04-11'
sub_test_start_date = '2016-04-11'
sub_test_end_date = '2016-04-16'
user_index, training_data, label = make_train_set(train_start_date,
train_end_date,
test_start_date,
test_end_date)
X_train, X_test, y_train, y_test = train_test_split(training_data,
label,
test_size=0.2,
random_state=0)
param = {
'n_estimators': 1200,
'max_depth': 3,
'subsample': 1.0,
'learning_rate': 0.01,
'min_samples_leaf': 1,
'random_state': 3,
'max_features': 0.8
}
clf = gbdt(param)
clf.fit(X_train, y_train)
sub_user_index, sub_trainning_date, sub_label = make_train_set(
sub_start_date, sub_end_date, sub_test_start_date,
sub_test_end_date) # use this data to see the offline score
test = sub_trainning_date.values
y = clf.predict(test)
pred = sub_user_index.copy()
y_true = get_labels_8(
sub_test_start_date,
sub_test_end_date) # during the test date, real label for cate 8
# y_true = sub_user_index.copy()
pred[
'label'] = y # add the new column which is the predict label for the test date
ans = []
for i in range(0, 30):
pred = sub_user_index.copy()
pred['label'] = y
pred = pred[pred.label >= i / 100]
# print(pred)
rep = report(pred, y_true)
print('%s : score:%s' % (i / 100, rep))
ans.append([i / 100, rep])
print('ans:%s' % ans)
threshold = sorted(ans, key=getKey, reverse=True)[0][0]
bestscore = sorted(ans, key=getKey, reverse=True)[0][1]
print('best threshold:%s' % threshold)
print('best score:%s' % bestscore)
def xgboost_report_submission():
train_start_date = '2016-03-08'
train_end_date = '2016-04-09'
result_start_date = '2016-04-09'
result_end_date = '2016-04-14'
valid_start_date = '2016-03-01'
valid_end_date = '2016-04-02'
valid_result_start_date = '2016-04-02'
valid_result_end_date = '2016-04-07'
test_start_date = '2016-03-15'
test_end_date = '2016-04-16'
user_index, training_data, label = make_train_set(train_start_date, train_end_date, result_start_date,
result_end_date)
x_train, x_test, y_train, y_test = train_test_split(training_data.values,
label.values, test_size=0.2, random_state=0)
dtrain = xgb.DMatrix(x_train, label=y_train)
dtest = xgb.DMatrix(x_test, label=y_test)
param = {'learning_rate': 0.1, 'n_estimators': 1000, 'max_depth': 3,
'min_child_weight': 5, 'gamma': 0, 'subsample': 1.0, 'colsample_bytree': 0.8,
'scale_pos_weight': 1, 'eta': 0.05, 'silent': 1, 'objective': 'binary:logistic'}
num_round = 283
param['nthread'] = 4
#param['eval_metric'] = "auc"
plst = param.items()
plst += [('eval_metric', 'logloss')]
evallist = [(dtest, 'eval'), (dtrain, 'train')]
bst = xgb.train(plst, dtrain, num_round, evallist)
# Report with validation set
valid_user_index, valid_trainning_date = make_test_set(valid_start_date, valid_end_date)
valid_trainning_date = xgb.DMatrix(valid_trainning_date.values)
pred_y = bst.predict(valid_trainning_date)
valid_pred = valid_user_index.copy()
valid_pred['label'] = pred_y
valid_pred = valid_pred[valid_pred['label'] >= 0.014]
valid_pred = valid_pred.sort_values('label', ascending=False).groupby('user_id').first().reset_index()
valid_true = get_true(valid_result_start_date, valid_result_end_date)
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.016]
print 0.016
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.018]
print 0.018
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.02]
print 0.02
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.022]
print 0.022
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.024]
print 0.024
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.026]
print 0.026
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.028]
print 0.028
report(valid_pred, valid_true)
valid_pred = valid_pred[valid_pred['label'] >= 0.03]
print 0.03
report(valid_pred, valid_true)