如何防止梯度提升机过拟合？

Question

如何防止梯度提升机过拟合？

6

我正在比较几个模型（梯度提升机、随机森林、逻辑回归、支持向量机、多层感知器和Keras神经网络）在一个多分类问题上的表现。我使用了嵌套交叉验证和网格搜索来运行我的模型，对实际数据和随机数据都进行了检验以检查是否过拟合。然而，对于梯度提升机，无论我如何改变数据或模型参数，它每次都能在随机数据上达到100％的准确率。是否有什么问题导致这种情况发生？

以下是我的代码：

dataset= pd.read_csv('data.csv')
data = dataset.drop(["gene"],1)
df = data.iloc[:,0:26]
df = df.fillna(0)
X = MinMaxScaler().fit_transform(df)

le = preprocessing.LabelEncoder()
encoded_value = le.fit_transform(["certain", "likely", "possible", "unlikely"])
Y = le.fit_transform(data["category"])

sm = SMOTE(random_state=100)
X_res, y_res = sm.fit_resample(X, Y)

seed = 7
logreg = LogisticRegression(penalty='l1', solver='liblinear',multi_class='auto')
LR_par= {'penalty':['l1'], 'C': [0.5, 1, 5, 10], 'max_iter':[100, 200, 500, 1000]}

rfc =RandomForestClassifier(n_estimators=500)
param_grid = {"max_depth": [3],
             "max_features": ["auto"],
              "min_samples_split": [2],
              "min_samples_leaf": [1],
              "bootstrap": [False],
              "criterion": ["entropy", "gini"]}


mlp = MLPClassifier(random_state=seed)
parameter_space = {'hidden_layer_sizes': [(50,50,50)],
     'activation': ['relu'],
     'solver': ['adam'],
     'max_iter': [10000],
     'alpha': [0.0001],
     'learning_rate': ['constant']}

gbm = GradientBoostingClassifier()
param = {"loss":["deviance"],
    "learning_rate": [0.001],
    "min_samples_split": [2],
    "min_samples_leaf": [1],
    "max_depth":[3],
    "max_features":["auto"],
    "criterion": ["friedman_mse"],
    "n_estimators":[50]
    }

svm = SVC(gamma="scale")
tuned_parameters = {'kernel':('linear', 'rbf'), 'C':(1,0.25,0.5,0.75)}

inner_cv = KFold(n_splits=10, shuffle=True, random_state=seed)

outer_cv = KFold(n_splits=10, shuffle=True, random_state=seed)


def baseline_model():

    model = Sequential()
    model.add(Dense(100, input_dim=X_res.shape[1], activation='relu')) #dense layers perform: output = activation(dot(input, kernel) + bias).
    model.add(Dropout(0.5))
    model.add(Dense(50, activation='relu')) #8 is the dim/ the number of hidden units (units are the kernel)
    model.add(Dense(4, activation='softmax'))

    model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])
    return model

models = []

models.append(('GBM', GridSearchCV(gbm, param, cv=inner_cv,iid=False, n_jobs=1)))
models.append(('RFC', GridSearchCV(rfc, param_grid, cv=inner_cv,iid=False, n_jobs=1)))
models.append(('LR', GridSearchCV(logreg, LR_par, cv=inner_cv, iid=False, n_jobs=1)))
models.append(('SVM', GridSearchCV(svm, tuned_parameters, cv=inner_cv, iid=False, n_jobs=1)))
models.append(('MLP', GridSearchCV(mlp, parameter_space, cv=inner_cv,iid=False, n_jobs=1)))
models.append(('Keras', KerasClassifier(build_fn=baseline_model, epochs=100, batch_size=50, verbose=0)))

results = []
names = []
scoring = 'accuracy'
X_train, X_test, Y_train, Y_test = train_test_split(X_res, y_res, test_size=0.2, random_state=0)


for name, model in models:
    nested_cv_results = model_selection.cross_val_score(model, X_res, y_res, cv=outer_cv, scoring=scoring)
    results.append(nested_cv_results)
    names.append(name)
    msg = "Nested CV Accuracy %s: %f (+/- %f )" % (name, nested_cv_results.mean()*100, nested_cv_results.std()*100)
    print(msg)
    model.fit(X_train, Y_train)
    print('Test set accuracy: {:.2f}'.format(model.score(X_test, Y_test)*100),  '%')

输出：

Nested CV Accuracy GBM: 90.952381 (+/- 2.776644 )
Test set accuracy: 90.48 %
Nested CV Accuracy RFC: 79.285714 (+/- 5.112122 )
Test set accuracy: 75.00 %
Nested CV Accuracy LR: 91.904762 (+/- 4.416009 )
Test set accuracy: 92.86 %
Nested CV Accuracy SVM: 94.285714 (+/- 3.563483 )
Test set accuracy: 96.43 %
Nested CV Accuracy MLP: 91.428571 (+/- 4.012452 )
Test set accuracy: 92.86 %

随机数据代码：

ran = np.random.randint(4, size=161)
random = np.random.normal(500, 100, size=(161,161))
rand = np.column_stack((random, ran))
print(rand.shape)
X1 = rand[:161]
Y1 = rand[:,-1]
print("Random data counts of label '1': {}".format(sum(ran==1)))
print("Random data counts of label '0': {}".format(sum(ran==0)))
print("Random data counts of label '2': {}".format(sum(ran==2)))
print("Random data counts of label '3': {}".format(sum(ran==3)))

for name, model in models:
    cv_results = model_selection.cross_val_score(model, X1, Y1,  cv=outer_cv, scoring=scoring)
    names.append(name)
    msg = "Random data CV %s: %f (+/- %f)" % (name, cv_results.mean()*100, cv_results.std()*100)
    print(msg)

随机数据输出：

Random data CV GBM: 100.000000 (+/- 0.000000)
Random data CV RFC: 62.941176 (+/- 15.306485)
Random data CV LR: 23.566176 (+/- 6.546699)
Random data CV SVM: 22.352941 (+/- 6.331220)
Random data CV MLP: 23.639706 (+/- 7.371392)
Random data CV Keras: 22.352941 (+/- 8.896451)

这个梯度提升分类器(GBM)的准确率已经达到了100%，无论我是否减少特征数量，更改网格搜索中的参数（我会放入多个参数，但运行时间可能长达数小时而没有结果，所以我现在暂时放弃了这个问题），即使我尝试二进制分类数据也是如此。

随机森林(RFC)的准确率也较高，为62%，我做错了什么吗？

我使用的数据主要是二进制特征，例如下面这样（并预测类别列）：

gene   Tissue    Druggable Eigenvalue CADDvalue Catalogpresence   Category
ACE      1           1         1          0           1            Certain
ABO      1           0         0          0           0            Likely
TP53     1           1         0          0           0            Possible

任何指导都将不胜感激。

- DN1

也许有一个有趣的相关问题：https://stats.stackexchange.com/questions/372676/theoretically-can-gradient-boosting-achieve-100-of-accuracy-in-an-arbitrary-dat - Eskapp

谢谢您分享这个，我会进一步研究的，因为我是初学者，但是乍一看，我是否正确地认为这意味着我的模型可能以某种方式使用无限深度？ - DN1

您可能会对以下内容感兴趣：梯度提升是否会过拟合，其中我重点关注了修改梯度提升算法中估计器数量的影响。 - RUser4512

1个回答

网页内容由stack overflow 提供, 点击上面的

可以查看英文原文，
原文链接

- sonia · Accepted Answer

一般来说，有几个参数可以用来减少过拟合。最容易理解的是增加min_samples_split和min_samples_leaf。将这些值设置得更高，将不允许模型记住如何正确识别单个数据或非常小的数据组。对于大型数据集（约1百万行），如果不是更高，我会将这些值设置在约50。您可以进行网格搜索以找到适合您特定数据的值。

您还可以使用subsample来减少过拟合，以及max_features。这些参数基本上不让您的模型查看某些数据，从而防止其记忆它们。