def preprocess_xy(self, label):
x, y = pc.get_filtered_x_y(self.feature_set_df, self.profile_df, label)
x = x.dropna(how='all', axis=0)
x = x.dropna(how='all', axis=1)
if x.isnull().any().any() or (x == np.inf).any().any() or (x == -np.inf).any().any():
x_imp = pc.fill_nan_features(x)
else:
x_imp = x
y_filtered = y[(map(int, x.columns.values))]
return x_imp, y_filtered
def compute_randomized_lr_score(data_set_df, user_info_df, label='gender'):
# print "\t\t\tfilling nan values..."
df_filtered, y_v = pc.get_filtered_x_y(data_set_df, user_info_df, label)
x = df_filtered.dropna(how='all')
x_imp = pc.fill_nan_features(x) if x.isnull().any().any() else x.values
clf = RandomizedLogisticRegression()
# print "\t\t\tfitting LR model..."
clf.fit(x_imp.T, y_v)
feature_importances = DataFrame(clf.scores_, index=df_filtered.index, columns=['importance'])
feature_importances.sort_values('importance', ascending=False, inplace=True, na_position='last')
return feature_importances
def compute(x):
x_imp = pc.fill_nan_features(x)
try:
m = ExtraTreesClassifier(n_estimators=n_est) if max_depth is None \
else ExtraTreesClassifier(n_estimators=n_est, max_depth=3)
print "\t\t\tfitting RF model..."
m.fit(x_imp.T, y_v)
# if len(feature_mics) > 1000:
# break
# print m.feature_importances_
for order, index in enumerate(x.index):
feature_importances.loc[index] = m.feature_importances_[order]
if float(order) % 10000 == 0 and order > 0:
print "\t\t\t%s features are done" % order
except ValueError as e:
# print "value error occurs during processing %r" % index
pass
def classify(data_set_df, user_info_df, feat_set_name, features=None, label='gender',
classifier=None, reg_param=1.0, selection=False, num_feat=20, sel_method='LR',
cv=10):
instance_num = len(data_set_df.columns)
df_filtered, y_v = pc.get_filtered_x_y(data_set_df, user_info_df, label)
x = df_filtered if features is None else df_filtered.loc[features]
x = x.dropna(how='all', axis=0)
x = x.dropna(how='all', axis=1)
if x.isnull().any().any() or (x == np.inf).any().any() or (x == -np.inf).any().any():
x_imp = pc.fill_nan_features(x)
# x_imp = dense_df.loc[x.index, x.columns]
else:
x_imp = x
y_filtered = y_v[(map(int, x.columns.values))]
clf = LogisticRegression(C=reg_param) if classifier is None else classifier
cv_num = min(len(y_filtered), cv)
score_mean = 0.0
miss_clf_rate = 1.0
if cv_num > 1 and len(y_filtered.unique()) > 1:
kf = KFold(y_filtered.shape[0], n_folds=cv_num, shuffle=True)
# skf = StratifiedKFold(y_filtered, n_folds=cv_num, shuffle=True)
fold = 0
result_str = ""
matrix_str = ""
for tr_index, te_index in kf:
fold += 1
x_train, x_test = x_imp.T.iloc[tr_index], x_imp.T.iloc[te_index]
y_train, y_test = y_filtered.iloc[tr_index], y_filtered.iloc[te_index]
if selection:
if sel_method == 'LR' or 'RF' in sel_method:
feat_index = fimp.feature_selection(x_train.T, user_info_df, num_feat,
method=sel_method, label=label)
else:
x_tr_df, x_te_df = x.T.iloc[tr_index].T, x.T.iloc[te_index].T
feat_index = fimp.feature_selection(x_tr_df, user_info_df, num_feat,
method=sel_method, label=label)
x_train = x_train.loc[:, feat_index].values
x_test = x_test.loc[:, feat_index].values
try:
clf.fit(x_train, y_train)
score = clf.score(x_test, y_test)
score_mean += score
result_str += "%s, %s, %s, %s, %s, %s, %s, %s, %s, %s\n" \
% (label, True if param.FILL_SUFFIX in feat_set_name else False,
True if param.SCALING_SUFFIX in feat_set_name else False, selection, 'LR',
reg_param, cv, fold, x_train.shape[1], score)
cf_mat = confusion_matrix(y_test, clf.predict(x_test),
labels=range(len(info.LABEL_CATEGORY[label])))
matrix_str += np.array_str(cf_mat) + "\n"
except ValueError:
pass
# traceback.print_exc()
# print i, "why error? skip!"
print result_str
file_name = "%s/new_%s.csv" % (param.EXPERIMENT_PATH, feat_set_name)
with open(file_name, mode='a') as f:
f.write(result_str)
file_name = "%s/new_%s_mat.csv" % (param.EXPERIMENT_PATH, feat_set_name)
with open(file_name, mode='a') as f:
f.write(matrix_str)
if fold > 0:
score_mean = score_mean / fold
miss_clf_rate = (float(instance_num - len(y_filtered)) / instance_num)
return score_mean, miss_clf_rate
