def main(path, delimiter, verbose):
# loading the dataframe
data_frame, all_headers = read_csv(path, delimiter, verbose)
# selecting headers of interest
headers = ['id',
#'hash_email',
#'hash_email_conversion',
#'hash_userid',
'rank',
'occurrences',
'lifetime',
'nb_days',
'nb_idtags',
'nb_idtags_site',
'nb_idtags_media',
#'click_rate',
'nb_purchases',
'last_time',
'nb_ips']
headers_to_drop = list(set(all_headers) - set(headers))
headers_to_scale = headers[:]
headers_to_scale.remove('id')
headers_to_scale.remove('rank')
# K-Fold cross-validation
nb_folds = 10
fold = KFolds(data_frame, nb_folds)
missranked_scores_train = []
missranked_scores_test = []
for k in range(nb_folds):
train, test = fold.get_fold(k)
train = train.sort_values(by='id')
test = test.sort_values(by='id')
# dropping not usefull columns
for drop in headers_to_drop:
train = train.drop(drop, 1)
test = test.drop(drop, 1)
# train set
train, mean, std = scaling(train, headers_to_scale)
train = train.reset_index(drop=True)
X = train[headers_to_scale].values
Y = train['rank']==1
# training Logistic regression model
linreg = LinearRegression(fit_intercept=True, normalize=False)
linreg.fit(X, Y)
# computing score on train set
Y_score_train = linreg.predict(X)
Y_predicted_train = compute_prediction(train, Y_score_train, verbose)
missranked_train, wellranked_train, total_train = compute_error(Y, Y_predicted_train)
missranked_scores_train.append(missranked_train/total_train)
# test set
test = scaling(test, headers_to_scale, mean, std)
test = test.reset_index(drop=True)
X_test = test[headers_to_scale].values
Y_test = test['rank'].values==1
# computing score on test set
Y_score_test = linreg.predict(X_test)
Y_predicted_test = compute_prediction(test, Y_score_test, verbose)
missranked_test, wellranked_test, total_test = compute_error(Y_test, Y_predicted_test)
missranked_scores_test.append(missranked_test/total_test)
# printing intermediate results
if verbose:
print('\n**** fold ', k, '****')
print('train set:')
print(' missranked =', round(missranked_train/total_train, 3))
print(' wellranked =', round(wellranked_train/total_train, 3))
print('test set:')
print(' missranked =', round(missranked_test/total_test, 3))
print(' wellranked =', round(wellranked_test/total_test, 3))
# printing final result
if verbose:
print('\n******** MEAN over all folds ********')
print('Train missranked = ', np.mean(missranked_scores_train))
print(' Test missranked = ', np.mean(missranked_scores_test))
def main(train_path, test_path):
delimiter = ';'
verbose = True
data_frame, all_headers = read_csv(train_path, delimiter, verbose)
delimiter_test = ','
verbose = True
data_frame_test, all_headers_test = read_csv(test_path, delimiter_test,
verbose)
headers = [
'id', 'rank', 'occurrences', 'lifetime', 'nb_days', 'nb_idtags',
'nb_idtags_site', 'nb_idtags_media', 'nb_purchases', 'last_time',
'nb_ips'
]
headers_test = [
'id', 'similarity', 'occurrences', 'lifetime', 'nb_days', 'nb_idtags',
'nb_idtags_site', 'nb_idtags_media', 'nb_purchases', 'last_time',
'nb_ips'
]
headers_to_drop = list(set(all_headers) - set(headers))
headers_to_drop_test = list(set(all_headers_test) - set(headers_test))
headers_to_scale = headers[:]
headers_to_scale.remove('id')
headers_to_scale.remove('rank')
headers_to_scale_test = headers_test[:]
headers_to_scale_test.remove('id')
headers_to_scale_test.remove('similarity')
train = data_frame.copy()
train = train.sort_values(by='id')
test = data_frame_test.copy()
test = test.sort_values(by='id')
for drop in headers_to_drop:
train = train.drop(drop, 1)
for drop in headers_to_drop_test:
test = test.drop(drop, 1)
train = train.reset_index(drop=True)
X = train[headers_to_scale].values
Y = train['rank'] == 1
similarity = np.array(test.loc[:, 'similarity'])
baseline = np.zeros(similarity.shape)
test = test.reset_index(drop=True)
test['arg_max'] = 0
old_id = test.loc[0, 'id']
rows = []
for row in range(test.shape[0]):
current_id = test.loc[row, 'id']
if current_id == old_id:
rows.append(row)
else:
sample = np.random.randint(min(rows), max(rows) + 1)
baseline[sample] = 1
old_id = current_id
m = test.loc[rows, 'similarity'].max()
test.loc[rows, 'arg_max'] = (test.loc[rows, 'similarity'] == m)
rows = []
rows.append(row)
if row == test.shape[0] - 1:
sample = np.random.randint(min(rows), max(rows) + 1)
baseline[sample] = 1
old_id = current_id
m = test.loc[rows, 'similarity'].max()
test.loc[rows, 'arg_max'] = (test.loc[rows, 'similarity'] == m)
rows = []
rows.append(row)
X_test = test[headers_to_scale].values
Y_test = test['arg_max']
# training model
regularization = 1e10
xgb_reg = GradientBoostingClassifier(loss='exponential',
n_estimators=50,
criterion='friedman_mse',
max_depth=5,
verbose=verbose)
xgb_reg.fit(X, Y)
# computing score on train set
Y_score_train = xgb_reg.predict_proba(X)[:, 1]
Y_predicted_train = compute_prediction(train, Y_score_train) #, verbose)
missranked_train, wellranked_train, total_train = compute_error(
Y, Y_predicted_train)
# printing intermediate results
print('train set:')
print(' missranked =', round(missranked_train / total_train, 3))
print(' wellranked =', round(wellranked_train / total_train, 3))
#
# computing score on test set
Y_score_test = xgb_reg.predict_proba(X_test)[:, 1]
Y_predicted_test = compute_prediction(test, Y_score_test) #, verbose)
missranked_test, wellranked_test, total_test = compute_error(
Y_test, Y_predicted_test)
missranked_test_baseline, wellranked_test_baseline, total_test_baseline = compute_error(
Y_test, baseline)
# printing intermediate results
print('test set:')
print(' missranked =', round(missranked_test / total_test, 3))
print(' wellranked =', round(wellranked_test / total_test, 3))
print('baseline prediction:')
print(' missranked =',
round(missranked_test_baseline / total_test_baseline, 3))
print(' wellranked =',
round(wellranked_test_baseline / total_test_baseline, 3))
# Les deux métriques du pdf
#df_2 = data_frame_test.copy()
old_id = test.loc[0, 'id']
rows = []
score_1 = 0
score_2 = 0
score_1_baseline = 0
score_2_baseline = 0
N = 0
for row in range(len(Y_predicted_test)):
current_id = test.loc[row, 'id']
if current_id == old_id:
rows.append(row)
if Y_predicted_test[row] == 1:
prediction = row
N += 1
if baseline[row] == 1:
prediction_baseline = row
else:
m = test.loc[rows, 'similarity'].max()
score_1 += m - test.loc[prediction, 'similarity']
score_2 += m * test.loc[prediction, 'similarity']
score_1_baseline += m - test.loc[prediction_baseline, 'similarity']
score_2_baseline += m * test.loc[prediction_baseline, 'similarity']
#print(str(N)+' : commande '+str(old_id)+' nb_lignes = '+str(len(rows))+' ; sim_max = '+str(m)+' ; sim_predicted = '+str(test.loc[prediction, 'similarity'])+' ; sim_baseline = '+str(test.loc[prediction_baseline, 'similarity']))
old_id = current_id
rows = []
rows.append(row)
if Y_predicted_test[row] == 1:
prediction = row
N += 1
if baseline[row] == 1:
prediction_baseline = row
if row == len(Y_predicted_test) - 1:
m = test.loc[rows, 'similarity'].max()
score_1 += m - test.loc[prediction, 'similarity']
score_2 += m * test.loc[prediction, 'similarity']
score_1_baseline += m - test.loc[prediction_baseline, 'similarity']
score_2_baseline += m * test.loc[prediction_baseline, 'similarity']
#print(str(N)+' : commande '+str(old_id)+' nb_lignes = '+str(len(rows))+' ; sim_max = '+str(m)+' ; sim_predicted = '+str(test.loc[prediction, 'similarity'])+' ; sim_baseline = '+str(test.loc[prediction_baseline, 'similarity']))
score_1 /= N
score_2 /= N
score_1_baseline /= N
score_2_baseline /= N
print('score_1 = ' + str(score_1))
print('score_2 = ' + str(score_2))
print('score_1_baseline = ' + str(score_1_baseline))
print('score_2_baseline = ' + str(score_2_baseline))