preprocess
# 通用的预处理框架
import pandas as pd
import numpy as np
import scipy as sp
# 文件读取
def read_csv_file(f, logging=False):
print("==========读取数据=========")
data = pd.read_csv(f)
if logging:
print(data.head(5))
print(f, "包含以下列")
print(data.columns.values)
print(data.describe())
print(data.info())
return data
Logistic Regression
# 通用的LogisticRegression框架
import pandas as pd
import numpy as np
from scipy import sparse
from sklearn.preprocessing import OneHotEncoder
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
# 1. load data
df_train = pd.DataFrame()
df_test = pd.DataFrame()
y_train = df_train['label'].values
# 2. process data
ss = StandardScaler()
# 3. feature engineering/encoding
# 3.1 For Labeled Feature
enc = OneHotEncoder()
feats = ["creativeID", "adID", "campaignID"]
for i, feat in enumerate(feats):
x_train = enc.fit_transform(df_train[feat].values.reshape(-1, 1))
x_test = enc.fit_transform(df_test[feat].values.reshape(-1, 1))
if i == 0:
X_train, X_test = x_train, x_test
else:
X_train, X_test = sparse.hstack((X_train, x_train)), sparse.hstack((X_test, x_test))
# 3.2 For Numerical Feature
# It must be a 2-D Data for StandardScalar, otherwise reshape(-1, len(feats)) is required
feats = ["price", "age"]
x_train = ss.fit_transform(df_train[feats].values)
x_test = ss.fit_transform(df_test[feats].values)
X_train, X_test = sparse.hstack((X_train, x_train)), sparse.hstack((X_test, x_test))
# model training
lr = LogisticRegression()
lr.fit(X_train, y_train)
proba_test = lr.predict_proba(X_test)[:, 1]
LightGBM
1. 二分类
import lightgbm as lgb
import pandas as pd
import numpy as np
import pickle
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split
print("Loading Data ... ")
# 导入数据
train_x, train_y, test_x = load_data()
# 用sklearn.cross_validation进行训练数据集划分，这里训练集和交叉验证集比例为7：3，可以自己根据需要设置
X, val_X, y, val_y = train_test_split(
train_x,
train_y,
test_size=0.05,
random_state=1,
stratify=train_y ## 这里保证分割后y的比例分布与原数据一致
)
X_train = X
y_train = y
X_test = val_X
y_test = val_y
# create dataset for lightgbm
lgb_train = lgb.Dataset(X_train, y_train)
lgb_eval = lgb.Dataset(X_test, y_test, reference=lgb_train)
# specify your configurations as a dict
params = {
'boosting_type': 'gbdt',
'objective': 'binary',
'metric': {'binary_logloss', 'auc'},
'num_leaves': 5,
'max_depth': 6,
'min_data_in_leaf': 450,
'learning_rate': 0.1,
'feature_fraction': 0.9,
'bagging_fraction': 0.95,
'bagging_freq': 5,
'lambda_l1': 1,
'lambda_l2': 0.001, # 越小l2正则程度越高
'min_gain_to_split': 0.2,
'verbose': 5,
'is_unbalance': True
}
# train
print('Start training...')
gbm = lgb.train(params,
lgb_train,
num_boost_round=10000,
valid_sets=lgb_eval,
early_stopping_rounds=500)
print('Start predicting...')
preds = gbm.predict(test_x, num_iteration=gbm.best_iteration) # 输出的是概率结果
# 导出结果
threshold = 0.5
for pred in preds:
result = 1 if pred > threshold else 0
# 导出特征重要性
importance = gbm.feature_importance()
names = gbm.feature_name()
with open('./feature_importance.txt', 'w+') as file:
for index, im in enumerate(importance):
string = names[index] + ', ' + str(im) + '\n'
file.write(string)
2.多分类
import lightgbm as lgb
import pandas as pd
import numpy as np
import pickle
from sklearn.metrics import roc_auc_score
from sklearn.model_selection import train_test_split
print("Loading Data ... ")
# 导入数据
train_x, train_y, test_x = load_data()
# 用sklearn.cross_validation进行训练数据集划分，这里训练集和交叉验证集比例为7：3，可以自己根据需要设置
X, val_X, y, val_y = train_test_split(
train_x,
train_y,
test_size=0.05,
random_state=1,
stratify=train_y ## 这里保证分割后y的比例分布与原数据一致
)
X_train = X
y_train = y
X_test = val_X
y_test = val_y
# create dataset for lightgbm
lgb_train = lgb.Dataset(X_train, y_train)
lgb_eval = lgb.Dataset(X_test, y_test, reference=lgb_train)
# specify your configurations as a dict
params = {
'boosting_type': 'gbdt',
'objective': 'multiclass',
'num_class': 9,
'metric': 'multi_error',
'num_leaves': 300,
'min_data_in_leaf': 100,
'learning_rate': 0.01,
'feature_fraction': 0.8,
'bagging_fraction': 0.8,
'bagging_freq': 5,
'lambda_l1': 0.4,
'lambda_l2': 0.5,
'min_gain_to_split': 0.2,
'verbose': 5,
'is_unbalance': True
}
# train
print('Start training...')
gbm = lgb.train(params,
lgb_train,
num_boost_round=10000,
valid_sets=lgb_eval,
early_stopping_rounds=500)
print('Start predicting...')
preds = gbm.predict(test_x, num_iteration=gbm.best_iteration) # 输出的是概率结果
# 导出结果
for pred in preds:
result = prediction = int(np.argmax(pred))
# 导出特征重要性
importance = gbm.feature_importance()
names = gbm.feature_name()
with open('./feature_importance.txt', 'w+') as file:
for index, im in enumerate(importance):
string = names[index] + ', ' + str(im) + '\n'
file.write(string)
XGBoost
1. 二分类
import numpy as np
import pandas as pd
import xgboost as xgb
import time
from sklearn.model_selection import StratifiedKFold
from sklearn.model_selection import train_test_split
train_x, train_y, test_x = load_data()
# 构建特征
# 用sklearn.cross_validation进行训练数据集划分，这里训练集和交叉验证集比例为7：3，可以自己根据需要设置
X, val_X, y, val_y = train_test_split(
train_x,
train_y,
test_size=0.01,
random_state=1,
stratify=train_y
)
# xgb矩阵赋值
xgb_val = xgb.DMatrix(val_X, label=val_y)
xgb_train = xgb.DMatrix(X, label=y)
xgb_test = xgb.DMatrix(test_x)
# xgboost模型 #####################
params = {
'booster': 'gbtree',
# 'objective': 'multi:softmax', # 多分类的问题、
# 'objective': 'multi:softprob', # 多分类概率
'objective': 'binary:logistic',
'eval_metric': 'logloss',
# 'num_class': 9, # 类别数，与 multisoftmax 并用
'gamma': 0.1, # 用于控制是否后剪枝的参数,越大越保守，一般0.1、0.2这样子。
'max_depth': 8, # 构建树的深度，越大越容易过拟合
'alpha': 0, # L1正则化系数
'lambda': 10, # 控制模型复杂度的权重值的L2正则化项参数，参数越大，模型越不容易过拟合。
'subsample': 0.7, # 随机采样训练样本
'colsample_bytree': 0.5, # 生成树时进行的列采样
'min_child_weight': 3,
# 这个参数默认是 1，是每个叶子里面 h 的和至少是多少，对正负样本不均衡时的 0-1 分类而言
# ，假设 h 在 0.01 附近，min_child_weight 为 1 意味着叶子节点中最少需要包含 100 个样本。
# 这个参数非常影响结果，控制叶子节点中二阶导的和的最小值，该参数值越小，越容易 overfitting。
'silent': 0, # 设置成1则没有运行信息输出，最好是设置为0.
'eta': 0.03, # 如同学习率
'seed': 1000,
'nthread': -1, # cpu 线程数
'missing': 1,
'scale_pos_weight': (np.sum(y==0)/np.sum(y==1)) # 用来处理正负样本不均衡的问题,通常取：sum(negative cases) / sum(positive cases)
# 'eval_metric': 'auc'
}
plst = list(params.items())
num_rounds = 2000 # 迭代次数
watchlist = [(xgb_train, 'train'), (xgb_val, 'val')]
# 交叉验证
result = xgb.cv(plst, xgb_train, num_boost_round=200, nfold=4, early_stopping_rounds=200, verbose_eval=True, folds=StratifiedKFold(n_splits=4).split(X, y))
# 训练模型并保存
# early_stopping_rounds 当设置的迭代次数较大时，early_stopping_rounds 可在一定的迭代次数内准确率没有提升就停止训练
model = xgb.train(plst, xgb_train, num_rounds, watchlist, early_stopping_rounds=200)
model.save_model('../data/model/xgb.model') # 用于存储训练出的模型
preds = model.predict(xgb_test)
# 导出结果
threshold = 0.5
for pred in preds:
result = 1 if pred > threshold else 0
处理正负样本不均匀的案例
# 计算正负样本比例
positive_num = df_train[df_train['label']==1].values.shape[0]
negative_num = df_train[df_train['label']==0].values.shape[0]
print(float(positive_num)/float(negative_num))
主要思路
- 手动调整正负样本比例
- 过采样 Over-Sampling
对训练集里面样本数量较少的类别（少数类）进行过采样，合成新的样本来缓解类不平衡，比如SMOTE算法
- 欠采样 Under-Sampling
- 将样本按比例一一组合进行训练，训练出多个弱分类器，最后进行集成
