models_cr = dict(knn=KNeighborsClassifier(p=8,
n_neighbors=35,
leaf_size=70,
algorithm='kd_tree'),
mlp=skorch_murder_bot(input_size=cr_input_size,
hidden_layers=np.array([
cr_input_size * 2,
cr_input_size // 2
])),
svm=LinearSVC(class_weight='balanced'),
baggingTrees=BaggingClassifier(
DecisionTreeClassifier(criterion='entropy',
ccp_alpha=0.005)),
boosting=GradientBoostingClassifier(ccp_alpha=0.005))
compare_models_all_metrics(models_cr, x_cr, y_cr, train_sizes)
models_sh = dict(knn=KNeighborsClassifier(p=8,
n_neighbors=35,
leaf_size=70,
algorithm='kd_tree'),
mlp=skorch_murder_bot(input_size=sh_input_size,
hidden_layers=np.array([
sh_input_size * 2,
sh_input_size // 2
])),
svm=LinearSVC(class_weight='balanced'),
baggingTrees=BaggingClassifier(
DecisionTreeClassifier(criterion='entropy',
ccp_alpha=0.005)),
boosting=GradientBoostingClassifier(ccp_alpha=0.005))
if __name__ == "__main__":
# Load the data
x_cr, y_cr = load_normalized_credit_fraud_numpy(
filepath='../data/creditcard.csv')
x_sh, y_sh = load_normalized_shopper_intention_numpy(
filepath='../data/online_shoppers_intention.csv')
train_sizes = np.linspace(0.3, 0.6, 4)
x_sh = x_sh.astype(np.float32)
input_size = x_sh.shape[1]
nets = {
'murderbot':
skorch_murder_bot(input_size=input_size,
hidden_layers=np.array(
[input_size * 2, input_size // 2])),
}
# compare_models_all_metrics(nets, x_sh, y_sh, train_sizes)
x_cr = x_cr.astype(np.float32)
y_cr = y_cr.astype(np.int)
input_size = x_cr.shape[1]
nets = {
'murderbot':
skorch_murder_bot(input_size=input_size,
hidden_layers=np.array(
[input_size * 2, input_size // 2])),
}
compare_models_all_metrics(nets, x_cr, y_cr, train_sizes)
DecisionTreeClassifier(criterion='entropy',
class_weight='balanced',
ccp_alpha=a)))
train_sizes = np.linspace(0.2, 1, 7)
# Show learning curve comparison of gini vs entropy
# TODO: the P-R and ROC curves aren't too useful, why?
# compare_models_all_metrics(models_criterion, x_cr, y_cr, train_sizes=train_sizes, title_prefix="Credit Fraud")
# Show learning curve comparison of balanced vs unbalanced
# compare_models_all_metrics(models_balancing, x_cr, y_cr, train_sizes=train_sizes, title_prefix="Credit Fraud")
compare_models_all_metrics(ensembles,
x_cr,
y_cr,
train_sizes=train_sizes,
title_prefix="Credit Fraud")
x_sub, x_val, y_sub, y_val = train_test_split(x_cr,
y_cr,
test_size=0.3,
random_state=0)
x_train, x_test, y_train, y_test = train_test_split(x_sub,
y_sub,
test_size=0.6,
random_state=0)
# Show how to determine an optimal ccpa value for post-pruning
# tree = DecisionTreeClassifier(criterion='entropy', class_weight='balanced')
# plot_nodes_vs_alpha(tree, x_train, y_train)
algorithm='kd_tree',
weights='distance',
n_jobs=-1)
train_sizes = np.linspace(0.1, 1, 10)
# Compare different n_neighbors values
clfs_neighbors = dict()
for n_neighbors in range(15, 56, 5):
clf = copy.deepcopy(base)
clf.n_neighbors = n_neighbors
clfs_neighbors['{}-nn'.format(n_neighbors)] = clf
compare_models_all_metrics(clfs_neighbors,
x_cr,
y_cr,
train_sizes=train_sizes,
title_prefix="Credit Fraud",
plot_learning_curve=False)
# Compare different leaf sizes
clfs_leaf_size = dict()
for leaf_size in range(5, 71, 10):
clf = copy.deepcopy(base)
clf.leaf_size = leaf_size
clfs_leaf_size['{}-leaf_size'.format(leaf_size)] = clf
compare_models_all_metrics(clfs_leaf_size,
x_cr,
y_cr,
train_sizes=train_sizes,
title_prefix="Credit Fraud",