LGB+XGB+CNN一般写法
阅读原文时间:2023年07月13日阅读:3

现在的比赛,想要拿到一个好的名次,就一定要进行模型融合,这里总结一下三种基础的模型:
- lightgbm:由于现在的比赛数据越来越大,想要获得一个比较高的预测精度,同时又要减少内存占用以及提升训练速度,lightgbm是一个非常不错的选择,其可达到与xgboost相似的预测效果。
- xgboost:在lightgbm出来之前,是打比赛的不二之选,现在由于需要做模型融合以提高预测精度,所以也需要使用到xgboost。
- ANN:得益于现在的计算机技术的高度发展,以及GPU性能的提高,还有Keras,tensorflow,pytorch等多重工具的使用,人工神经网络也可以作为最后模型融合的子模型之一,可以有效地提升最终的预测结果。

下面附上使用三个函数的Python代码,可以直接运行。(参考:https://blog.csdn.net/meyh0x5vdtk48p2/article/details/78816334)

LGB

def LGB_predict(train_x,train_y,test_x,res,index):
print("LGB test")
clf = lgb.LGBMClassifier(
boosting_type='gbdt', num_leaves=31, reg_alpha=0.0, reg_lambda=1,
max_depth=-1, n_estimators=5000, objective='binary',
subsample=0.7, colsample_bytree=0.7, subsample_freq=1,
learning_rate=0.05, min_child_weight=50, random_state=2018, n_jobs=-1
)
clf.fit(train_x, train_y, eval_set=[(train_x, train_y)], eval_metric='auc',early_stopping_rounds=100)
res['score'+str(index)] = clf.predict_proba(test_x)[:,1]
res['score'+str(index)] = res['score'+str(index)].apply(lambda x: float('%.6f' % x))
print(str(index)+' predict finish!')
gc.collect()
res=res.reset_index(drop=True)
return res['score'+str(index)]

XGB

def XGB_predict(train_x,train_y,val_X,val_Y,test_x,res):
print("XGB test")
# create dataset for lightgbm

xgb\_val = xgb.DMatrix(val\_X, label=val\_Y)  
xgb\_train = xgb.DMatrix(X\_train, label=y\_train)  
xgb\_test = xgb.DMatrix(test\_x)  
# specify your configurations as a dict  
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': 'auc',  
          # '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 = 5000 # 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)  
res\['score'\] = model.predict(xgb\_test)  
res\['score'\] = res\['score'\].apply(lambda x: float('%.6f' % x))  
return res

CNN

imp = Imputer(missing_values='NaN', strategy='mean', axis=0)
X_train = imp.fit_transform(X_train)
sc = StandardScaler(with_mean=False)
sc.fit(X_train)
X_train = sc.transform(X_train)
val_X = sc.transform(val_X)
X_test = sc.transform(X_test)

ann_scale = 1

from keras.layers import Embedding

model = Sequential()

model.add(Embedding(X_train.shape[1] + 1,
EMBEDDING_DIM,
input_length=MAX_SEQUENCE_LENGTH))
#model.add(Dense(int(256 / ann_scale), input_shape=(X_train.shape[1],)))
model.add(Dense(int(256 / ann_scale)))
model.add(Activation('tanh'))
model.add(Dropout(0.3))
model.add(Dense(int(512 / ann_scale)))
model.add(Activation('relu'))
model.add(Dropout(0.3))
model.add(Dense(int(512 / ann_scale)))
model.add(Activation('tanh'))
model.add(Dropout(0.3))
model.add(Dense(int(256 / ann_scale)))
model.add(Activation('linear'))
model.add(Dense(1))
model.add(Activation('sigmoid'))

For a multi-class classification problem

model.summary()

class_weight1 = class_weight.compute_class_weight('balanced',
np.unique(y),
y)

#-----------------------------------------------------------------------------------------------------------------------------------------------------

AUC for a binary classifier

def auc(y_true, y_pred):
ptas = tf.stack([binary_PTA(y_true,y_pred,k) for k in np.linspace(0, 1, 1000)],axis=0)
pfas = tf.stack([binary_PFA(y_true,y_pred,k) for k in np.linspace(0, 1, 1000)],axis=0)
pfas = tf.concat([tf.ones((1,)) ,pfas],axis=0)
binSizes = -(pfas[1:]-pfas[:-1])
s = ptas*binSizes
return K.sum(s, axis=0)

PFA, prob false alert for binary classifier

def binary_PFA(y_true, y_pred, threshold=K.variable(value=0.5)):
y_pred = K.cast(y_pred >= threshold, 'float32')
# N = total number of negative labels
N = K.sum(1 - y_true)
# FP = total number of false alerts, alerts from the negative class labels
FP = K.sum(y_pred - y_pred * y_true)
return FP/N

P_TA prob true alerts for binary classifier

def binary_PTA(y_true, y_pred, threshold=K.variable(value=0.5)):
y_pred = K.cast(y_pred >= threshold, 'float32')
# P = total number of positive labels
P = K.sum(y_true)
# TP = total number of correct alerts, alerts from the positive class labels
TP = K.sum(y_pred * y_true)
return TP/P
#-----------------------------------------------------------------------------------------------------------------------------------------------------

model.compile(loss='binary_crossentropy',
optimizer='rmsprop',

metrics=['accuracy'],

          metrics=\[auc\])  

epochs = 100
model.fit(X_train, y, epochs=epochs, batch_size=2000,
validation_data=(val_X, val_y), shuffle=True,
class_weight = class_weight1)