def main():
# Prepare data
df = pd.read_csv('data/creditcard.csv')
df = df.drop(['Time'], axis=1)
sc = StandardScaler()
df.iloc[:, :-1] = sc.fit_transform(df.iloc[:, :-1])
X = df.iloc[:, :-1]
y = df.iloc[:, -1]
skf = StratifiedKFold(n_splits=3, random_state=123)
fpr, tpr, roc_auc = dict(), dict(), dict()
# Logistic Regression
y_true, y_pred_proba = [], []
for train_index, test_index in skf.split(X, y):
clf = LogisticRegression()
clf.fit(X.iloc[train_index, :], y[train_index])
y_true.append(y[test_index])
y_pred_proba.append(clf.predict_proba(X.iloc[test_index, :])[:, 1])
y_pred_proba = np.concatenate(y_pred_proba)
y_true = np.concatenate(y_true)
fpr['lr'], tpr['lr'], _ = roc_curve(y_true, y_pred_proba)
roc_auc['lr'] = np.round(roc_auc_score(y_true, y_pred_proba), 3)
# GBM
y_true, y_pred_proba = [], []
for train_index, test_index in skf.split(X, y):
clf = GradientBoostingClassifier()
clf.fit(X.iloc[train_index, :], y[train_index])
y_true.append(y[test_index])
y_pred_proba.append(clf.predict_proba(X.iloc[test_index, :])[:, 1])
y_pred_proba = np.concatenate(y_pred_proba)
y_true = np.concatenate(y_true)
fpr['gbm'], tpr['gbm'], _ = roc_curve(y_true, y_pred_proba)
roc_auc['gbm'] = np.round(roc_auc_score(y_true, y_pred_proba), 3)
# LOF
y_true, y_pred_proba = [], []
for train_index, test_index in skf.split(X, y):
X_train_unsupervised = X.iloc[train_index, :][y[train_index] == 0]
clf = LocalOutlierFactor()
clf.fit(X_train_unsupervised)
decision_function = clf._decision_function(X.iloc[test_index, :])
y_true.append(y[test_index])
y_pred_proba.append(1 - np.interp(decision_function, \
(decision_function.min(),
decision_function.max()), (0, 1)))
y_pred_proba = np.concatenate(y_pred_proba)
y_true = np.concatenate(y_true)
fpr['lof'], tpr['lof'], _ = roc_curve(y_true, y_pred_proba)
roc_auc['lof'] = np.round(roc_auc_score(y_true, y_pred_proba), 3)
# Isolation Forest
y_true, y_pred_proba = [], []
for train_index, test_index in skf.split(X, y):
X_train_unsupervised = X.iloc[train_index, :][y[train_index] == 0]
clf = IsolationForest()
clf.fit(X_train_unsupervised)
decision_function = clf.decision_function(X.iloc[test_index, :])
y_true.append(y[test_index])
y_pred_proba.append(1 - np.interp(decision_function, \
(decision_function.min(),
decision_function.max()), (0, 1)))
y_pred_proba = np.concatenate(y_pred_proba)
y_true = np.concatenate(y_true)
fpr['if'], tpr['if'], _ = roc_curve(y_true, y_pred_proba)
roc_auc['if'] = np.round(roc_auc_score(y_true, y_pred_proba))
# Autoencoder unsupervised
y_true, y_pred_proba = [], []
for train_index, test_index in skf.split(X, y):
X_train_unsupervised = X.iloc[train_index, :][y[train_index] == 0]
input_dim = X_train_unsupervised.shape[1]
clf = autoencoder.autoencoder_unsupervised(input_dim=input_dim)
clf.fit(X_train_unsupervised,
X_train_unsupervised,
batch_size=50,
epochs=3,
verbose=1)
y_true.append(y[test_index])
X_test_pred = clf.predict(X.iloc[test_index, :])
y_pred_proba.append(autoencoder.\
reconstruction_error(X.iloc[test_index, :], X_test_pred))
y_pred_proba = np.concatenate(y_pred_proba)
y_true = np.concatenate(y_true)
fpr['ae_unsupervised'], tpr['ae_unsupervised'], _ = roc_curve(
y_true, y_pred_proba)
roc_auc['ae_unsupervised'] = np.round(roc_auc_score(y_true, y_pred_proba),
3)
# Autoencoder supervised
y_true, y_pred_proba = [], []
for train_index, test_index in skf.split(X, y):
input_dim = X.iloc[train_index, :].shape[1]
y_train = pd.concat([X.iloc[train_index, :], y[train_index]], axis=1)
clf = autoencoder.autoencoder_supervised(input_dim=input_dim)
clf.fit(X.iloc[train_index, :],
y_train,
batch_size=50,
epochs=3,
verbose=0)
y_true.append(y[test_index])
X_test_pred = clf.predict(X.iloc[test_index, :])
y_pred_proba.append(autoencoder.\
reconstruction_error(X.iloc[test_index, :], X_test_pred))
y_pred_proba = np.concatenate(y_pred_proba)
y_true = np.concatenate(y_true)
fpr['ae_supervised'], tpr['ae_supervised'], _ = roc_curve(
y_true, y_pred_proba)
roc_auc['ae_supervised'] = np.round(roc_auc_score(y_true, y_pred_proba), 3)
# Plot ROC AUC curves
colors = ['blue', 'green', 'red', 'orange', 'purple', 'darkblue', 'black']
for i, model in enumerate(fpr.keys()):
print(str(model))
print(roc_auc[model])
plt.plot(fpr[model], tpr[model], color=colors[i])
plt.legend(roc_auc.items())
plt.title('ROC AUC curves all models')
plt.show()
plt.savefig('results/plots/roc_curves_all_models.png')
