def train_cv(label):
"""
cross validation to fine tune hypeparameters
:param label:
:return:
"""
df_train = pd.read_csv('../input/train.csv', encoding='utf-8')
clf = SVC(C=0.8, kernel='rbf', class_weight='balanced')
# feature composition
df_handcraft_train = pd.read_csv('../input/train_features.csv')[features]
tfidf_unigram_train = train_tfidf_unigram_features()
tfidf_bigram_train = train_tfidf_bigram_features()
tfidf_char_train = train_tfidf_char_features()
X_train = hstack((df_handcraft_train, tfidf_unigram_train,
tfidf_bigram_train, tfidf_char_train))
y_train = df_train[label]
grid_params = {
'C': np.arange(0.6, 1.0, 0.1),
'kernel': ('linear', 'rbf', 'poly', 'sigmoid'),
}
'''re-sample the data set'''
X_train_resampled, y_train_resampled = resample(X_train, y_train)
grid_clf = GridSearchCV(estimator=clf,
param_grid=grid_params,
verbose=1,
scoring='roc_auc',
cv=5)
grid_clf.fit(X_train_resampled, y_train_resampled)
return grid_clf
def train(label):
"""
train model process
:param label: target label list
:return:
"""
clf = SVC(C=0.8, kernel='rbf', class_weight='balanced', probability=True)
df_train = pd.read_csv('../input/train_clean.csv', encoding='utf-8')
'''feature composition'''
df_handcraft_train = pd.read_csv('../input/train_features.csv',
encoding='utf-8')[features].as_matrix()
# tfidf_unigram_train = train_tfidf_unigram_features()
tfidf_bigram_train = train_tfidf_bigram_features()
# tfidf_char_train = train_tfidf_char_features()
X_train = features_merge(df_handcraft_train, tfidf_bigram_train)
y_train = df_train[label]
# '''re-sample the data set'''
# X_train_resampled, y_train_resampled = resample(X_train, y_train)
# '''feature selection'''
# model = SelectFromModel(estimator=clf)
# X_train_resampled = model.transform(X_train_resampled)
# '''train test split'''
# X_train, X_valid, y_train, y_valid = train_test_split(X_train_resampled, y_train_resampled, test_size=0.2,
# random_state=2)
clf.fit(X_train, y_train)
# y_predict = clf.predict(X_valid)
# print(label + ' roc auc score is ' + str(roc_auc_score(y_valid, y_predict)))
# clf.fit(X_train_resampled, y_train_resampled)
return clf
def train_cv(label):
"""
cross validation to fine tune hypeparameters
:param label:
:return:
"""
df_train = pd.read_csv('../input/train.csv', encoding='utf-8')
clf = RandomForestClassifier(n_estimators=1000,
class_weight='balanced',
verbose=1)
# feature composition
df_handcraft_train = pd.read_csv('../input/train_features.csv')[features]
tfidf_unigram_train = train_tfidf_unigram_features()
tfidf_bigram_train = train_tfidf_bigram_features()
tfidf_char_train = train_tfidf_char_features()
X_train = hstack(df_handcraft_train, tfidf_unigram_train,
tfidf_bigram_train, tfidf_char_train)
y_train = df_train[label]
grid_params = {
'n_estimators': range(1000, 4000, 100),
'min_samples_leaf': range(2, 10, 1)
}
grid_clf = GridSearchCV(estimator=clf,
param_grid=grid_params,
verbose=1,
scoring='roc_auc',
cv=5)
grid_clf.fit(X_train, y_train)
return grid_clf
def train_cv(label):
"""
better params:
- C=4
- solver='lbfgs'
gird search to tune hypeparameters
:param label: target label
:return: lr model with special parameters
"""
# label_weight = df_train.shape[0]/df_train[df_train[label]==1].shape[0]
lr = LogisticRegression(class_weight='balanced',
solver='sag',
random_state=22,
verbose=1,
max_iter=6000,
C=4)
# lr.set_params(class_weight={label:label_weight})
# lr.set_params(class_weight='balanced')
df_train = pd.read_csv('../input/train.csv', encoding='utf-8')
'''feature composition'''
df_handcraft_train = pd.read_csv('../input/train_features.csv',
encoding='utf-8')[features].as_matrix()
tfidf_unigram_train = train_tfidf_unigram_features()
tfidf_bigram_train = train_tfidf_bigram_features()
tfidf_char_train = train_tfidf_char_features()
print(df_handcraft_train.shape)
# print(tfidf_train.shape)
X_train = features_merge(df_handcraft_train, tfidf_unigram_train,
tfidf_bigram_train, tfidf_char_train)
# X_train = np.concatenate((df_handcraft_train,df_tfidf_train),axis=1)
# df_train = pd.read_csv('../input/train.csv', encoding='utf-8')
y_train = df_train[label]
params = {
# 'penalty': ('l1', 'l2'),
# 'C': np.arange(1, 5, 1),
'solver': ('liblinear', 'sag', 'saga', 'newton-cg', 'lbfgs'),
# 'solver': ('sag', 'saga', 'newton-cg'),
# 'max_iter': range(6000, 7000, 100)
}
clf = GridSearchCV(estimator=lr,
param_grid=params,
scoring='roc_auc',
cv=5,
verbose=1)
clf.fit(X_train, y_train)
return clf
def train(label):
df_train = pd.read_csv('../input/train.csv', encoding='utf-8')
clf = XGBClassifier(
learning_rate=0.1,
n_estimators=2000,
max_depth=4,
silent=False,
# scale_pos_weight=_toxic_count / _clean_count,
colsample_bytree=0.8,
colsample_bylevel=0.6,
gamma=2,
objective='binary:logistic')
'''feature composition'''
df_handcraft_train = pd.read_csv('../input/train_features.csv',
encoding='utf-8')[features].as_matrix()
tfidf_unigram_train = train_tfidf_unigram_features()
tfidf_bigram_train = train_tfidf_bigram_features()
tfidf_char_train = train_tfidf_char_features()
X_train = features_merge(df_handcraft_train, tfidf_unigram_train,
tfidf_bigram_train, tfidf_char_train)
y_train = df_train['label']
# '''resample the data set'''
# X_train_resampled, y_train_resampled = resample(X_train, y_train)
'''feature selection'''
model = SelectFromModel(estimator=clf)
X_train = model.transform(X_train)
'''train test split'''
X_train, X_valid, y_train, y_valid = train_test_split(X_train,
y_train,
test_size=0.2,
random_state=2)
clf.fit(X_train,
y_train,
eval_metric='auc',
eval_set=[(X_valid, y_valid)],
early_stopping_rounds=20)
y_predict = clf.predict(X_valid)
print(label + ' roc auc score is ' +
str(roc_auc_score(y_valid, y_predict)))
# clf.fit(X_train_resampled, y_train_resampled)
return clf
def train_cv(label):
"""
cross validation training process
:param label:
:return: list clfs
"""
df_train = pd.read_csv('../input/train.csv', encoding='utf-8')
num_fold = 5
skf = StratifiedKFold(n_splits=num_fold, shuffle=True)
count = 0
clf_list = []
'''feature composition'''
df_handcraft_train = pd.read_csv('../input/train_features.csv',
encoding='utf-8')[features].as_matrix()
tfidf_unigram_train = train_tfidf_unigram_features()
tfidf_bigram_train = train_tfidf_bigram_features()
tfidf_char_train = train_tfidf_char_features()
word_embedding_train = pd.read_csv(
'../feature_engineering/word_embedding/w2v_train_embedding.csv',
encoding='utf-8')
X_train = features_merge(df_handcraft_train, tfidf_unigram_train,
tfidf_bigram_train, tfidf_char_train,
word_embedding_train)
# X_train = features_merge(df_handcraft_train, tfidf_unigram_train, tfidf_bigram_train, tfidf_char_train)
y_train = df_train[label]
for idx_train, idx_valid in skf.split(y_train, y_train):
print("fitting fold " + str(count))
clf = _train(label,
x_train=X_train,
y_train=y_train,
idx_train=idx_train,
idx_valid=idx_valid,
index=count)
count += 1
clf_list.append(clf)
return clf_list
def train_grid_search(label):
"""
better params:
- colsample_bytree=0.8
- colsample_bylevel=0.6
- gamma:2
- max_depth=4
cross validation to fine tune hypeparameters
:param label:
:return:
"""
df_train = pd.read_csv('../input/train.csv', encoding='utf-8')
_toxic_count = len(df_train[df_train.max(axis=1) == 0])
print(_toxic_count)
_clean_count = len(df_train) - _toxic_count
clf = XGBClassifier(
learning_rate=0.1,
n_estimators=2000,
max_depth=4,
silent=False,
# scale_pos_weight=_toxic_count / _clean_count,
colsample_bytree=0.8,
colsample_bylevel=0.6,
gamma=2,
objective='binary:logistic')
# feature composition
df_handcraft_train = pd.read_csv(
'../feature_engineering/statistics/statics_train.csv')
tfidf_unigram_train = train_tfidf_unigram_features()
tfidf_bigram_train = train_tfidf_bigram_features()
tfidf_char_train = train_tfidf_char_features()
word_embedding_train = pd.read_csv(
'../feature_engineering/word_embedding/w2v_train_embedding.csv',
encoding='utf-8')
# print(word_embedding_train.shape)
X_train = features_merge(df_handcraft_train, tfidf_unigram_train,
tfidf_bigram_train, tfidf_char_train,
word_embedding_train)
# X_train = features_merge(df_handcraft_train, tfidf_unigram_train, tfidf_bigram_train, tfidf_char_train)
# X_train = features_merge(df_handcraft_train, tfidf_unigram_train, tfidf_char_train)
y_train = df_train[label]
grid_params = {
# 'learning_rate': np.arange(0.08, 0.2, 0.01),
# 'n_estimators': range(1000, 4000, 100),
# 'gamma': range(0, 5, 1),
# 'max_depth': range(4, 5, 1),
# 'colsample_bytree': np.arange(0.6, 1, 0.1),
'colsample_bylevel': np.arange(0.5, 1.0, 0.1)
}
# '''resample the data set'''
# X_train_resampled, y_train_resampled = resample(X_train, y_train)
X_train, X_valid, y_train, y_valid = train_test_split(X_train,
y_train,
test_size=0.2,
random_state=2)
print('grid search construct')
grid_clf = GridSearchCV(estimator=clf,
param_grid=grid_params,
verbose=1,
scoring='roc_auc',
cv=5)
grid_clf.fit(X=X_train,
y=y_train,
eval_metric='auc',
eval_set=[(X_valid, y_valid)],
early_stopping_rounds=50)
return grid_clf
