def random_forest_churn(X_train, y_train, X_test, y_test):
fea_path = "conf/u_plan_train_conf"
col_names = []
with open(fea_path, 'r') as file:
for line in file:
line = line.strip('\n')
col_names.append(line)
# model construction
estimator_RF = RandomForestClassifier(n_estimators=116, max_depth=18, max_features=10, random_state=0, n_jobs=-11)
# model training
model_RF = estimator_RF.fit(X_train, y_train)
# model prediction
predict_RF = model_RF.predict(X_test)
y_predict_proba = model_RF.predict_proba(X_test)
model_validation.model_valid(y_test, predict_RF)
accu_scr, prec_scr, rec_scr, f1_scr = model_validation.model_valid(y_test, predict_RF)
print("RandomForest模型准确率 : " + str(accu_scr))
print("RandomForest模型精确率 : " + str(prec_scr))
print("RandomForest模型召回率 : " + str(rec_scr))
print("RandomForest模型F1 score : " + str(f1_scr))
# generate roc
fpr, tpr, thresholds = roc_curve(y_test, y_predict_proba[:, 1])
roc_auc = auc(fpr, tpr)
# get RF feature importance
importance_score_list = []
for i in range(len(model_RF.feature_importances_)):
importance_score_list.append([model_RF.feature_importances_[i], col_names[i]])
importance_score_list = sorted(importance_score_list, key=lambda x: x[0], reverse=True)
# get RF importance features
print("=" * 50)
print("xgboost feature importance")
for i in range(len(importance_score_list)):
print(importance_score_list[i][1] + '\t' + str(importance_score_list[i][0]))
print("=" * 50)
# # 得分
# score_RF_train = model_RF.score(X_train, y_train)
# score_RF_test = model_RF.score(X_test, y_test)
# # 特征重要性
# fig, ax = plt.subplots(figsize=(7, 5))
# ax.bar(col_names, model_RF.feature_importances_)
# ax.set_title("Feature Importances")
# fig.savefig('pic/RF_feature_importance.png')
return fpr, tpr, roc_auc
def lightGBM_churn(X_train, y_train, X_test, y_test):
# model construction
lgb_model = lgb.LGBMClassifier(boosting_type='gbdt',
num_leaves=63,
max_depth=-1,
learning_rate=0.1,
n_estimators=200,
max_bin=255,
subsample_for_bin=2000,
objective=None,
min_split_gain=0.0,
min_child_weight=2,
scale_pos_weight=2,
min_child_samples=20,
subsample=0.8,
subsample_freq=1,
colsample_bytree=0.8,
reg_alpha=1,
reg_lambda=1,
random_state=12,
n_jobs=8,
silent=True)
# model training
lgb_churn_model = lgb_model.fit(X_train,
y_train,
eval_metric='auc',
eval_set=[(X_test, y_test)],
early_stopping_rounds=100)
y_pred = lgb_churn_model.predict(X_test)
y_pred_proba = lgb_churn_model.predict_proba(X_test)
accu_scr, prec_scr, rec_scr, f1_scr = model_validation.model_valid(
y_test, y_pred)
print("lightGBM模型准确率 : " + str(accu_scr))
print("lightGBM模型精确率 : " + str(prec_scr))
print("lightGBM模型召回率 : " + str(rec_scr))
print("lightGBM模型F1 score : " + str(f1_scr))
# generate roc
fpr, tpr, thresholds = roc_curve(y_test, y_pred_proba[:, 1])
roc_auc = auc(fpr, tpr)
return fpr, tpr, roc_auc
def xgboost_churn(X_train, y_train, X_test, y_test):
# load conf
name = "xgboost_churn"
fea_path = "conf/u_plan_train_conf"
col_names = ['y']
with open(fea_path, 'r') as file:
for line in file:
line = line.strip('\n')
col_names.append(line)
real_fea_names = col_names[1:]
X_df = pd.DataFrame(X_train, columns=real_fea_names)
# model training
dtrain = xgb.DMatrix(X_train, y_train)
lmda = 1.0
params = {
"objective": "binary:logistic",
"max_depth": 6,
"eta": 0.3,
"gamma": 1,
"colsample_bytree": 0.8,
"min_child_weight": 5,
"subsample": 0.8
}
best_iteration = 150
bst = xgb.train(params, dtrain, best_iteration)
# model validation
dtest = xgb.DMatrix(X_test)
y_pred_proba = bst.predict(dtest)
y_pred = np.array([1 if item >= 0.5 else 0 for item in y_pred_proba])
accu_scr, prec_scr, rec_scr, f1_scr = model_validation.model_valid(
y_test, y_pred)
print("xgboost模型准确率 : " + str(accu_scr))
print("xgboost模型精确率 : " + str(prec_scr))
print("xgboost模型召回率 : " + str(rec_scr))
print("xgboost模型F1 score : " + str(f1_scr))
# generate roc
fpr, tpr, thresholds = roc_curve(y_test, y_pred_proba)
roc_auc = auc(fpr, tpr)
# get importance score and rank
importance_score_dict = bst.get_score(importance_type="gain")
importance_score_list = []
for k, v in importance_score_dict.items():
importance_score_list.append((v, int(k[1:])))
importance_score_list = sorted(importance_score_list,
key=lambda x: x[0],
reverse=True)
# get xgboost importance features
print("=" * 50)
print("xgboost feature importance")
for i in range(len(importance_score_list)):
print(real_fea_names[importance_score_list[i][1]] + '\t' +
str(importance_score_list[i][0]))
print("=" * 50)
top_20_f = []
for i in range(20):
top_20_f.append(real_fea_names[importance_score_list[i][1]])
# # explainer model training
# params = {"objective": "binary:logistic", 'silent': 1, 'eval_metric': 'auc', 'base_score': 0.5, "lambda": lmda}
# bst = xgb.train(params, dtrain, best_iteration)
#
# # calculate xgboost probability
# tree_lst = xgb_exp.model2table(bst, lmda=lmda)
# leaf_lsts = bst.predict(dtrain, pred_leaf=True)
# fea_logit = [[] for _ in range(len(real_fea_names))]
# for i, leaf_lst in enumerate(leaf_lsts):
# dist = xgb_exp.logit_contribution(tree_lst, leaf_lst)
# for idx in range(len(real_fea_names)):
# if 'f' + str(idx) in dist:
# fea_logit[idx].append(-dist['f' + str(idx)])
# else:
# fea_logit[idx].append(0)
#
# # explainer plot
# fig = plt.figure(figsize=(30, 30))
# fig.suptitle('Feature User Churn')
# for i in range(len(top_20_f)):
# fea = top_20_f[i]
# idx = real_fea_names.index(fea)
# fea_data = X_df[fea].values
# ax = fig.add_subplot(5, 4, i + 1)
# ax.set_xscale('log')
# ax.set_xlim(min([i for i in fea_data if i > 0]), max(fea_data))
# ax.scatter(fea_data, fea_logit[idx], s=0.1)
# ax.hlines(0, min([i for i in fea_data if i > 0]), max(fea_data), linewidth=0.5)
# ax.set_title(top_20_f[i])
# plt.subplots_adjust(wspace=0.3, hspace=0.3)
# plt.savefig('pic/' + name + '_' + str(best_iteration) + '.png', dpi=800)
return fpr, tpr, roc_auc