8 Credit Card Fraud
8.1 Load: Credit Card Fraud
- Source: Kaggle
Python
loc = 'data/Y_04_CCR'
loc_csv = loc + '.csv'
loc_bin = loc + '.feather'
if(not os.path.exists(loc_bin)):
#Read | Relocate | Rename | Factor | Levels | ID | Save |
pp = pd.read_csv(loc_csv)
if(False): pp.drop(columns = ['Time'], inplace = True) #Drop Columns
qq = list(pp.columns)
qq.insert(0, qq.pop(qq.index('Amount'))) #Relocate list element
qq.insert(0, qq.pop(qq.index('Class')))
pp = pp.reindex(columns = qq)
pp.columns.values[[np.arange(3, 3+9)]] = \
['V0' + str(i) for i in np.arange(1, 1+9)] #Rename V1 to V01
pp.rename(columns ={'Class': 'is_Found'}, inplace = True)
pp['is_Found'] = pp['is_Found'].astype('category')
pp['is_Found'] = pp['is_Found'].cat.rename_categories(['No', 'Yes'])
pp.insert(0, 'ID', range(1, 1+len(pp)))
pyarrow.feather.write_feather(pp, loc_bin) #Data: Credit Card Fraud
else:
pp = pyarrow.feather.read_feather(loc_bin)
y_ccr = pp.copy()
pp.shape
## (284807, 32)
qq = list(pp.columns) #Column Names
if(False): list(qq[0:5]) + list(qq[-3:len(qq)])
print('\n'.join([', '.join(qq[i:i+8]) for i in range(0,len(qq),8)]))
## ID, is_Found, Amount, Time, V01, V02, V03, V04
## V05, V06, V07, V08, V09, V10, V11, V12
## V13, V14, V15, V16, V17, V18, V19, V20
## V21, V22, V23, V24, V25, V26, V27, V28
qq = pp['is_Found'].value_counts()
qq = qq.to_frame(qq.index.name)
qq.index.name = None
qq['PCT'] = 100 * qq['is_Found'] / qq['is_Found'].sum() #Count & PCT
print(qq)
## is_Found PCT
## No 284315 99.827251
## Yes 492 0.172749
if(False): pp.info(memory_usage = False)
if(False): pp['Amount'].describe()
if(False): pp['Time'].max() #(2 Days) 2*24*60*60 - 88.3 Train & Test
Python
Python
# Split Train and Test with similar proportion of Response
y_trn_x_id, y_tst_x_id, y_trn_y_id, y_tst_y_id = \
sklearn.model_selection.train_test_split(y_ccr_X, y_ccr_Y,
test_size = 0.2, random_state = 3)
y_trn_y = y_trn_y_id['is_Found']
y_tst_y = y_tst_y_id['is_Found']
pp = [i for i in list(y_trn_x_id.columns) if i not in ['ID']]
y_trn_x = y_trn_x_id[pp]
y_tst_x = y_tst_x_id[pp]
if(True): y_trn_x.shape, y_tst_x.shape, y_trn_y.shape, y_tst_y.shape
## ((227845, 30), (56962, 30), (227845,), (56962,))
pp = y_trn_y.value_counts().to_frame()
pp['PCT'] = 100 * pp / pp.sum()
pp
## count PCT
## is_Found
## No 227453 99.827953
## Yes 392 0.172047
qq = y_tst_y.value_counts().to_frame()
qq['PCT'] = 100 * qq / qq.sum()
qq
## count PCT
## is_Found
## No 56862 99.824444
## Yes 100 0.1755568.4 Random Forest
Python
loc = 'data/Y_04_RFC.sav'
if(not os.path.exists(loc)):
# Fit
y_rfc = sklearn.ensemble.RandomForestClassifier(n_estimators = 100,
criterion = 'gini', random_state = 3, n_jobs = -1, verbose = 0)
y_rfc.fit(y_trn_x, y_trn_y)
joblib.dump(y_rfc, loc)
else:
y_rfc = joblib.load(loc)
list(y_rfc.classes_) #Classes Labels
## ['No', 'Yes']
y_rfc.n_features_in_ #Count Features
## 30
y_rfc_xs = list(y_rfc.feature_names_in_) #Features
print('\n'.join([', '.join(y_rfc_xs[i:i+8]) for i in range(0,len(y_rfc_xs),8)]))
## Amount, Time, V01, V02, V03, V04, V05, V06
## V07, V08, V09, V10, V11, V12, V13, V14
## V15, V16, V17, V18, V19, V20, V21, V22
## V23, V24, V25, V26, V27, V28Python
pp = [ round(i, 3) for i in y_rfc.feature_importances_ ]
#print('\n'.join([', '.join(map(str, pp[i:i+8])) for i in range(0,len(pp),8)]))
qq = pd.DataFrame(dict(Features = y_rfc_xs, Importance = pp))
qq.sort_values('Importance', ascending = False).head()
## Features Importance
## 18 V17 0.152
## 15 V14 0.150
## 13 V12 0.125
## 11 V10 0.094
## 17 V16 0.071Python
print(round(y_rfc.score(y_trn_x, y_trn_y), 5)) #Train
## 1.0
print(round(y_rfc.score(y_tst_x, y_tst_y), 5)) #Test
## 0.99942
# Confusion Matrix
pd.crosstab(y_tst_y, y_rfc_pred,
rownames = ['Actual'], colnames = ['Predicted'], margins = True)
## Predicted No Yes All
## Actual
## No 56855 7 56862
## Yes 26 74 100
## All 56881 81 56962
pp = sklearn.metrics.confusion_matrix(y_tst_y, y_rfc_pred,
labels = list(y_rfc.classes_))
if(False): print(pp)
with np.printoptions(precision = 5, suppress = True):
print(100 * pp/np.sum(pp)) #Percent
## [[99.81216 0.01229]
## [ 0.04564 0.12991]]
tn, fp, fn, tp = pp.ravel()
(tn, fp, fn, tp)
## (56855, 7, 26, 74)
pp = y_tst_y.values.astype('object')
pp.size
## 56962
y_unique, y_counts = np.unique(pp, return_counts = True)
print(np.asarray((y_unique, y_counts)).T)
## [['No' 56862]
## ['Yes' 100]]
qq = y_rfc_pred
qq.size
## 56962
y_unique, y_counts = np.unique(qq, return_counts = True)
print(np.asarray((y_unique, y_counts)).T)
## [['No' 56881]
## ['Yes' 81]]
print('Accuracy (Avoid): ', sklearn.metrics.accuracy_score(
y_tst_y, y_rfc_pred))
## Accuracy (Avoid): 0.999420666409185
print('Precision: ', sklearn.metrics.precision_score(
y_tst_y, y_rfc_pred, pos_label = 'Yes')) #tp/(tp+fp)
## Precision: 0.9135802469135802
print('Recall (Sensitivity) (More Important): ', sklearn.metrics.recall_score(
y_tst_y, y_rfc_pred, pos_label = 'Yes')) #tp/(tp+fn)
## Recall (Sensitivity) (More Important): 0.74
print('F1 (Avoid): ', sklearn.metrics.f1_score(
y_tst_y, y_rfc_pred, pos_label = 'Yes')) #2*tp/(2*tp+fn+fp)
## F1 (Avoid): 0.81767955801104978.5 Confusion Matrix
- Precision makes sure we do not spot good transactions as fraudulent in our problem.
- Recall assures we do not predict fraudulent transactions as all good.
R
# Confusion Matrix
# Actual Positive Negative
# Predicted
# Positive TP FP (I) Precision = TP /(TP+FP)
# Negative FN (II) TN NPV = TN /(TN+FN)
# Sensitivity Specificity Accuracy
# = TP /(TP+FN) = TN /(TN+FP) = (TP+TN) /N8.6 Undersampling
- To be done only on the Training Data, not on the Test Data
- Near Miss refers to a group of undersampling strategies that pick samples based on the distance between majority and minority class instances.
Python
#def q_get_best_model(estimator, params, kf = kf, label = 'Yes',
# train_X = y_trn_x, train_Y = y_trn_y, test_X = y_tst_x, test_Y = y_tst_y,
# is_grid_search = True, sampling = NearMiss(), scoring='f1', n_jobs = -1):
# if sampling is None:
# # make the pipeline of only the estimator
# pipeline = make_pipeline(estimator)
# else:
# # make the pipeline of over or undersampling and estimator
# pipeline = make_pipeline(sampling, estimator)
# # get the estimator name
# estimator_name = estimator.__class__.__name__.lower()
# # construct the parameters for grid/random search cv
# new_params = {f'{estimator_name}__{key}': params[key] for key in params}
# if is_grid_search:
# # grid search instead of randomized search
# search = sklearn.model_selection.GridSearchCV(pipeline,
# param_grid = new_params, cv = kf,
# return_train_score = True, n_jobs = n_jobs, verbose = 0)
# else:
# # randomized search
# search = sklearn.model_selection.RandomizedSearchCV(pipeline,
# param_distributions = new_params, cv = kf, scoring = scoring,
# return_train_score = True, n_jobs = n_jobs, verbose = 0)
# # fit the model
# search.fit(train_X, train_Y)
# cv_score = cross_val_score(search, train_X, train_Y,
# scoring = scoring, cv = kf)
# # make predictions on the test data
# y_pred = search.best_estimator_.named_steps[estimator_name].predict(test_X)
# # calculate the metrics: recall, accuracy, F1 score, etc.
# recall = sklearn.metrics.recall_score(test_Y, y_pred, pos_label = label)
# precision = sklearn.metrics.precision_score(test_Y, y_pred, pos_label = label)
#
# print('Hi') #xxxx
#
# accuracy = sklearn.metrics.accuracy_score(test_Y, y_pred)
# f1 = sklearn.metrics.f1_score(test_Y, y_pred, pos_label = label)
# y_proba = search.best_estimator_.named_steps[estimator_name].predict_proba(test_X)[::, 1]
# fpr, tpr, _ = roc_curve(test_Y, y_proba)
# auc = roc_auc_score(test_Y, y_proba)
# # return the best estimator along with the metrics
# return({
# 'best_estimator': search.best_estimator_,
# 'estimator_name': estimator_name,
# 'cv_score': cv_score,
# 'recall': recall,
# 'precision': precision,
# 'accuracy': accuracy,
# 'f1_score': f1,
# 'fpr': fpr,
# 'tpr': tpr,
# 'auc': auc,
# })
#8.7 Model
Python
if(False):
# Cross Validation Framework
from sklearn.model_selection import StratifiedKFold
from sklearn.model_selection import GridSearchCV, cross_val_score, RandomizedSearchCV
#kf = StratifiedKFold(n_splits = 5, random_state = None, shuffle = False)
# Imbalance
from imblearn.pipeline import make_pipeline
from imblearn.under_sampling import NearMiss
from imblearn.over_sampling import SMOTE
# Metrics
from sklearn.metrics import roc_curve, roc_auc_score
# Classifiers
from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
from sklearn.neighbors import KNeighborsClassifier
from sklearn.tree import DecisionTreeClassifier8.8 Tuning
Python
# Number of trees in random forest
#n_estimators = np.linspace(100, 3000, int((3000-100)/200) + 1, dtype=int)
n_estimators = np.array([100])
# Number of features to consider at every split
max_features = ['sqrt'] #['auto', 'sqrt']
# Maximum number of levels in tree
max_depth = [1, 2, 3]
# Minimum number of samples required to split a node
# min_samples_split = [int(x) for x in np.linspace(start = 2, stop = 10, num = 9)]
min_samples_split = [2, 3]
# Minimum number of samples required at each leaf node
min_samples_leaf = [1, 2]
# Method of selecting samples for training each tree
bootstrap = [True] #[True, False]
# Criterion
criterion = ['gini'] #['gini', 'entropy']
random_grid = {'n_estimators': n_estimators,
# 'max_features': max_features,
'max_depth': max_depth,
'min_samples_split': min_samples_split,
'min_samples_leaf': min_samples_leaf,
'bootstrap': bootstrap,
'criterion': criterion}8.9 Small Test Case
Python
y_rfc_base = sklearn.ensemble.RandomForestClassifier()
if(False):
y_rfc_cv_pre = sklearn.model_selection.RandomizedSearchCV(
estimator = y_rfc_base, param_distributions = random_grid,
n_iter = 3, cv = 2, verbose = 0, random_state = 3, n_jobs = -1)
y_rfc_cv_pre.fit(y_trn_x, y_trn_y)
print(y_rfc_cv_pre.best_params_)R
if(TRUE) py_config() #Python Configuration
## python: C:/Softwares/Python/Python312/python.exe
## libpython: C:/Softwares/Python/Python312/python312.dll
## pythonhome: C:/Softwares/Python/Python312
## version: 3.12.1 (tags/v3.12.1:2305ca5, Dec 7 2023, 22:03:25) [MSC v.1937 64 bit (AMD64)]
## Architecture: 64bit
## numpy: C:/Softwares/Python/Python312/Lib/site-packages/numpy
## numpy_version: 1.26.3
##
## NOTE: Python version was forced by use_python() function
if(FALSE) q_url[ , 'URL'] #List of URL of this Page
if(FALSE) q_() #R Objects of this Page excluding 'q_*'