import xgboost as xgb
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import sklearn.metrics as metrics
import joblib
from utils.dataprocess import preProc, preProcTest, toJson
from sklearn.model_selection import train_test_split, GridSearchCV
# xgb原生接口
dataTrain = pd.read_csv(r'C:\Users\ZY\Desktop\ML\VI_train.csv')
xTrain, yTrain, scaler = preProc(dataTrain, toNumpy=True)
# weight=反例数量/正例数量
weight = (yTrain.shape[0] - sum(yTrain)) * 1.0 / sum(yTrain)
xTrain, xValidation, yTrain, yValidation = train_test_split(xTrain, yTrain, test_size=0.2)
dtrain = xgb.DMatrix(xTrain, label=yTrain)
dvalidation = xgb.DMatrix(xValidation, label=yValidation)
params = {
# xgboost宏观特征参数
'booster': 'gbtree',
'nthread': 5, # 线程数
'silent': 0, # 为1时，静默开启
# booster参数
'eta': 0.1, # learning rate 通过减少每一步的权重，提高鲁棒性
'gamma': 0.1, # 节点分裂所需要的最小损失函数下降值
'max_depth': 9, # 最大树高，限制过拟合
'lambda': 2, # 权重的L2正则项
'alpha': 1, # 权重的L1正则项
'''
for i in range(self.layer_num):
x = self.relu[i](self.bns[i](self.hiddens[i](x)))
x = self.predict(x)
return x
# 读入数据
dataTrain = pd.read_csv(r'C:\Users\ZY\Desktop\ML\VI_train.csv')
weight_negative = sum(dataTrain['Response']) * 1.0 / dataTrain.shape[0]
# 由于类别不平衡，采用权重解决，分别设置负类和正类的权重
weights = [weight_negative,
1 - weight_negative] # [0.12293666666666667, 0.8770633333333333]
# pytorch要求权重输入为tensor
weights = torch.from_numpy(np.array(weights)).type(torch.FloatTensor)
xTrain, yTrain, scaler = preProc(dataTrain, toTensor=True)
torch.save(xTrain, 'xTrain.pt')
torch.save(yTrain, 'yTrain.pt')
xTrain, xValidation, yTrain, yValidation = train_test_split(xTrain,
yTrain,
test_size=0.2,
random_state=1)
torch_dataset = Data.TensorDataset(xTrain, yTrain)
loader = Data.DataLoader(
dataset=torch_dataset,
batch_size=BATCH_SIZE,
shuffle=True,
)
net = Net(14, 15, 2, 3)
print(net)
from sklearn.svm import SVC
import numpy as np
import pandas as pd
import joblib
import time
from utils.dataprocess import preProc, preProcTest, toJson
import sklearn.metrics as metrics
t1 = time.time()
# 训练数据处理
dataTrain = pd.read_csv(r'C:\Users\ZY\Desktop\ML\VI_train.csv')
xTrain, yTrain, scaler = preProc(dataTrain)
# 建立模型
max_iter = 100000000
# model里加上class_weight='balanced',等价于正负例分别乘以权重sum(负例)、sum(正例)
# fit里有参数sample_weight,为每个sample赋上权重，是长度等于sample数量的array
# 这两个作用相同，只使用一个
model = SVC(C=1.0, kernel='rbf', gamma='auto', tol=0.2, cache_size=1024, class_weight='balanced', max_iter=max_iter)
model.fit(xTrain, yTrain, sample_weight=None)
score = model.score(xTrain, yTrain)
print('Score:', score)
pred_y = model.predict(xTrain)
fscore = metrics.f1_score(yTrain, pred_y)
print('Fvalue:', fscore)
equal1count = sum(pred_y == 1)
print('预测结果为1的数量:', equal1count)
print(pred_y)
joblib.dump(model, 'carInsurancePredSVM.model')
print('迭代次数', max_iter, '耗时:', time.time() - t1)