python imblearn 处理样本不平衡，svm验证

参考地址：https://blog.youkuaiyun.com/tonydz0523/article/details/84325823

1.样本分布不均的解决方法：

1.1 .过采样 通过增加分类中样本较少的类别的采样数量来实现平衡，最直接的方法是简单复制小样本数据时加入随机噪声、干扰数据等。

1.2 .欠采样(下采样) 通过减少分类中多数类样本的数量来实现样本均衡，最直接的方法是随机去掉一些多数类样本来减小多数类的规模，缺点是会丢失多数类中的一些重要信息。

1.3 .设置权重 对不同样本数量的类别赋予不同的权重（通常会设置为与样本量成反比 n samples / (n class * n bincount(y) )

1.4 .集成思想 每次生成训练集时，使用全量小样本，同时从大样本量中随机抽取等量数据量，反复多次会得到多个训练模型。最后在应用时，采用组合（例如投票、加权投票等）产生分类预测结果。

2.python处理

2.1 生成数据集

from collections import Counter
from sklearn.datasets import  make_classification
x,y = make_classification(n_samples=100000,n_features=10,n_informative=2,n_redundant=2, n_repeated=0, n_classes=3, n_clusters_per_class=1, weights=[0.89,0.1,0.01], class_sep=0.3,random_state=2020)

Counter(y)

Counter({0: 88486, 1: 10211, 2: 1303})

2.2 RandomOverSample(过采样)

使用RandomOverSampler从少数类的样本中进行随机采样来增加新的样本使各个分类均衡

from imblearn.over_sampling import RandomOverSampler
x_resampled,y_resampled = RandomOverSampler(random_state=0).fit_sample(x,y)

Counter(y_resampled)

Counter({0: 88486, 1: 88486, 2: 88486})

2.3 SMOTE(过采样)

对于少数类样本a, 随机选择一个最近邻的样本b, 然后从a与b的连线上随机选取一个点c作为新的少数类样本

from imblearn.over_sampling import SMOTE
x_resampled,y_resampled = SMOTE().fit_sample(x,y)

Counter(y_resampled)

Counter({0: 88486, 1: 88486, 2: 88486})

2.4 ADASYM(过采样)

ADASYN: 关注的是在那些基于K最近邻分类器被错误分类的原始样本附近生成新的少数类样本

from imblearn.over_sampling import ADASYN
x_resampled,y_resampled = ADASYN().fit_sample(x,y)

Counter(y_resampled)

Counter({0: 88486, 1: 88198, 2: 88138})

2.5 RandomUnderSampler(下采样)

andomUnderSampler函数是一种快速并十分简单的方式来平衡各个类别的数据: 随机选取数据的子集.

from imblearn.under_sampling import RandomUnderSampler
x_resampled,y_resampled = RandomUnderSampler().fit_sample(x,y)

Counter(y_resampled)

Counter({0: 1303, 1: 1303, 2: 1303})

2.6 设置权重(sklearn SVM举例)

class_weight : {dict, 'balanced'}, optional
    Set the parameter C of class i to class_weight[i]*C for
    SVC. If not given, all classes are supposed to have
    weight one.
    The "balanced" mode uses the values of y to automatically adjust
    weights inversely proportional to class frequencies in the input data
    as ``n_samples / (n_classes * np.bincount(y))``

    
def __init__(self, C=1.0, kernel='rbf', degree=3, gamma='scale',
             coef0=0.0, shrinking=True, probability=False,
             tol=1e-3, cache_size=200, class_weight=None,
             verbose=False, max_iter=-1, decision_function_shape='ovr',
             break_ties=False,
             random_state=None):
    
estimator = svm.SVC(max_iter=1,probability=True,class_weight='balanced',random_state=1234) ##此处还可以采用网格寻参

3. 测试样本不均衡对svm的影响

from sklearn import svm
clf = svm.SVC(class_weight='balanced',max_iter=10000)

from collections import Counter
from sklearn.datasets import  make_classification
x,y = make_classification(n_samples=10000,n_features=10,n_informative=2,n_redundant=2, n_repeated=0, n_classes=2, n_clusters_per_class=1, weights=[0.999,0.001], class_sep=0.3,random_state=2020)

Counter(y)

Counter({0: 9945, 1: 55})

from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler

x_stand = StandardScaler().fit_transform(x)

x_train,x_test,y_train,y_test = train_test_split(x_stand,y,test_size=0.2)

clf.fit(x_train,y_train)
pred = clf.predict(x_test)

pred

array([0, 1, 0, ..., 0, 0, 0])

y_test

array([0, 0, 0, ..., 0, 0, 0])

from sklearn.metrics import accuracy_score
accuracy_score(pred,y_test)

0.911

from imblearn.over_sampling import RandomOverSampler
x_resampled,y_resampled = RandomOverSampler(random_state=0).fit_sample(x,y)

Counter(y_resampled)

Counter({0: 9945, 1: 9945})

x_stand = StandardScaler().fit_transform(x_resampled)

x_train,x_test,y_train,y_test = train_test_split(x_stand,y_resampled,test_size=0.2)

clf.fit(x_train,y_train)
pred = clf.predict(x_test)

pred

array([0, 1, 0, ..., 1, 1, 0])

accuracy_score(pred,y_test)

0.9595274007038713