def one_fold(schema, train, test, s_option, nt):
pred_a = np.zeros(len(test))
pred_c = np.zeros(len(test))
bag_cnt = 0
alpha_idx = 0
while True:
# bagging
newdata = sampling(train, s_option)
# base tree
tree = dt.create_decision_tree(newdata, schema, 1.0, DEPTH)
pred = dt.apply_rules(test, schema, tree)
pred_c = pred_c + pred
# alpha variation
pred_down, alpha_cnt_down = alpha_variation(schema, newdata, test,
tree, False)
pred_up, alpha_cnt_up = alpha_variation(schema, newdata, test,
tree, True)
pred_a = pred_a + ((pred_down + pred_up + pred) /
(alpha_cnt_down + alpha_cnt_up + 1.0))
alpha_idx = alpha_idx + alpha_cnt_down + alpha_cnt_up
bag_cnt = bag_cnt + 1
if bag_cnt > nt:
break
pred_a = pred_a / (bag_cnt)
pred_c = pred_c / (bag_cnt)
label = test[:,-1]
roc_a = st.auc(pred_a, label)
roc_c = st.auc(pred_c, label)
return roc_a, roc_c, (roc_a/roc_c), (float(alpha_idx)/nt)
def leat_ai_raw(schema, train, test, s_option, nt, lift, z_beta):
cov_c45 = 0.0
cov_leat = 0.0
data = np.vstack((train,test))
base_prob = dt.laplace_smoothing(data)
obj_prob = lift * base_prob
pred = np.zeros(len(data))
alpha_list = [-1.0, -0.75, -0.5, -0.25, 0.0, 0.25, 0.5, 1.0, 1.5, 1.75, 2.0,2.25, 2.5, 2.75, 3.0]
# base tree
output = []
nt = 1
for alpha in alpha_list:
tree = dt.create_decision_tree(data, schema, alpha, -1,
True, obj_prob, z_beta)
pred_new = dt.apply_rules(data, schema, tree)
pred_added = pred + pred_new
cov_new = float(np.sum(pred_added > 0))/len(data)
cov_orig = float(np.sum(pred > 0))/len(data)
if cov_new > cov_orig:
pred = pred_added
output.append([nt,cov_new])
nt = nt + 1
print nt, cov_new
return output
def correlation(schema, train, test):
pred_a = np.zeros(len(test))
pred_c = np.zeros(len(test))
bag_cnt = 0
alpha_idx = 0
corr_c45 = []
corr_beat = []
# bagging
base_data = sampling(train, "None")
# base tree
tree = dt.create_decision_tree(base_data, schema, 1.0, DEPTH)
base_pred = dt.apply_rules(test, schema, tree)
base_pred = base_pred - np.mean(base_pred)
base_cov = np.sqrt(np.sum(base_pred * base_pred))
for ii in range(10):
newdata = sampling(train, "Normal")
tree = dt.create_decision_tree(newdata, schema, 1.0, DEPTH)
pred = dt.apply_rules(test, schema, tree)
pred_c = pred - np.mean(pred)
cov_c = np.sqrt(np.sum(pred_c*pred_c))
corr_c45.append(np.dot(base_pred,pred_c)/base_cov/cov_c)
# alpha variation
pred_down, alpha_cnt_down = alpha_variation(schema, newdata, test,
tree, False)
pred_up, alpha_cnt_up = alpha_variation(schema, newdata, test,
tree, True)
if alpha_cnt_down > 0:
pred_a = pred_down
pred_a = pred_a - np.mean(pred_a)
cov_a = np.sqrt(np.sum(pred_a*pred_a))
corr_beat.append(np.dot(base_pred,pred_a)/base_cov/cov_a)
if alpha_cnt_up > 0:
pred_a = pred_up
pred_a = pred_a - np.mean(pred_a)
cov_a = np.sqrt(np.sum(pred_a*pred_a))
corr_beat.append(np.dot(base_pred,pred_a)/base_cov/cov_a)
if alpha_cnt_up==0 and alpha_cnt_down==0:
corr_beat.append(np.dot(base_pred,pred_c/base_cov/cov_c))
return corr_beat, corr_c45
def leat(schema, train, test, s_option, nt, lift, z_beta):
cov_c45 = 0.0
cov_leat = 0.0
data = np.vstack((train,test))
base_prob = dt.laplace_smoothing(data)
obj_prob = lift * base_prob
pred_a = np.zeros(len(data))
pred_c = np.zeros(len(data))
bag_cnt = 0
alpha_list = [-1.0, -0.75, -0.5,-0.25, 0.0, 0.25, 0.5, 1.5, 1.75, 2.0,2.25, 2.5, 2.75, 3.0]
while True:
# bagging
newdata = sampling(data, s_option)
# base tree
tree = dt.create_decision_tree(newdata, schema, 1.0, -1,
True, obj_prob, z_beta)
pred = dt.apply_rules(data, schema, tree)
pred_c = pred_c + pred
pred_a = pred_a + pred
for alpha in alpha_list:
tree = dt.create_decision_tree(newdata, schema, alpha, -1,
True, obj_prob, z_beta)
pred = dt.apply_rules(data, schema, tree)
pred_a = pred_a + pred
bag_cnt = bag_cnt + 1
if bag_cnt > nt:
break
cov_c45 = float(np.sum(pred_c > 0))/len(data)
cov_leat = float(np.sum(pred_a > 0))/len(data)
return cov_c45, cov_leat
def alpha_variation(schema, train, test, base_tree, direction):
alpha_cnt = 0
pred = np.zeros(len(test))
alpha = 1.0
alpha_tree = base_tree
while True:
alpha = select_alpha(alpha_tree, train, schema, alpha, direction)
if alpha != 1.0:
alpha_tree = dt.create_decision_tree(train, schema, alpha, DEPTH)
pred = pred + dt.apply_rules(test, schema, alpha_tree)
alpha_cnt = alpha_cnt + 1
else:
break
return pred, alpha_cnt