原文：http://blog.csdn.net/hero_fantao/article/details/35784773

不平衡学习方法

 

机器学习中样本不平衡问题大致分为两方面：

（1）类别中样本比率不平衡，但是几个类别的样本都足够多；

（2）类别中某类样本较少。

对第二个问题，其实不是我们重点，因为样本不足的话，覆盖空间是很小，如果特征足够多的话，这种数据对模型学习的价值也不大，所以，对这个问题，好的方法只能是找尽量多的小类样本来覆盖样本空间。

现在主要讨论第一个问题。

 

一: 采样方法

1. 随机重采样(random oversampling):

  样本不平衡时候，对小类样本就行随机重采样，以达到平衡。这种方法只是对小类样本进行简单的拷贝，缺点是容易over-fit，比如在决策树分类的时候，很有可能一个终端叶子节点的样本都是一个样本的拷贝而已，扩展性不足，这可能会提高模型训练的精度，但是对未知测试样本的预测可能是很差的。

   

2. 随机欠采样(random oversampling)：

     样本不平衡时候，对大类样本就行随机欠采样，就是取部分大类样本，以达到平衡。欠采样的问题是对样本减少可能会缺失样本空间中重要数据，降低准确性。

 

3. Synthetic Sampling with Data Generation

  对小类样本进行近似数据样本生成。对小类样本计算KNN，找出K个相近样本，根据K近邻样本于当前样本的距离，生成新的样本。

 

 

 这种方法突破了原有的简单的重复采样的方法，通过创建新的小样本，丰富了小样本的样本空间，弥补了小样本样本空间不足的问题。缺点是它对所有的小类样本都计算相同的KNN。试想下对于那些和大类样本有明显的区分度的小样本，对于这些产生多余的样本价值不大。

 

4. Adaptive Synthetic Sampling

  Adaptive Synthetic Sampling是一种修正方法，他试图增加小样本中和大类样本比较相近的样本sampling。

方法如下：

 

 

二 代价学习方法

一是从样本角度来看，尽量做到样本平衡，然后来用模型的学习。还有种就是通过设置不同样本误判的代价，比如设置小样本误判的代价大一些。个人的感觉，这种方法和一中重采样的效果差不多，牺牲一个换取另外一个。个人觉得一种好的方法是，正负样本不平衡时候，每次选取一部分大类样本和全部小样本，尽量平衡，训练一个模型。重复以上操作，训练得到若干模型，把这些模型做个voting，获得最终预测结果，可以效仿Adaboost，对每个模型进行加权。其实，voting的方法就能达到很不多的效果。

 

 

参考文献：

[1] He H, Garcia E A. Learning from imbalanced data[J]. Knowledge and Data Engineering, IEEE Transactions on, 2009, 21(9): 1263-1284.

[2] https://github.com/fmfn/UnbalancedDataset(2014/12/07 分享的一个模块)

 

附上Adaptive Synthetic Sampling源码：

 

               '''''         Created on 2014/03/09                  @author: dylan         '''          from sklearn.neighbors import NearestNeighbors          import numpy as np          import random                                        def get_class_count(y, minorityclasslabel = 1):              minorityclasslabel_count = len(np.where(y == minorityclasslabel)[0])              maxclasslabel_count = len(np.where(y == (1 - minorityclasslabel))[0])                            return maxclasslabel_count, minorityclasslabel_count                                      # @param: X The datapoints e.g.: [f1, f2, ... ,fn]          # @param: y the classlabels e.g: [0,1,1,1,0,...,Cn]          # @param ms: The amount of samples in the minority group          # @param ml: The amount of samples in the majority group          # @return: the G value, which indicates how many samples should be generated in total, this can be tuned with beta          def getG(ml, ms, beta):              return (ml-ms)*beta                              # @param: X The datapoints e.g.: [f1, f2, ... ,fn]          # @param: y the classlabels e.g: [0,1,1,1,0,...,Cn]          # @param: minorityclass: The minority class          # @param: K: The amount of neighbours for Knn          # @return: rlist: List of r values          def getRis(X,y,indicesMinority,minorityclasslabel,K):                                ymin = np.array(y)[indicesMinority]              Xmin = np.array(X)[indicesMinority]              neigh = NearestNeighbors(n_neighbors= K)              neigh.fit(X)                            rlist = [0]*len(ymin)              normalizedrlist = [0]*len(ymin)                            for i in xrange(len(ymin)):                  indices = neigh.kneighbors(Xmin[i],K,False)[0]          #         print'y[indices] == (1 - minorityclasslabel):'          #         print y[indices]          #         print len(np.where(y[indices] == ( 1- minorityclasslabel))[0])                  rlist[i] = len(np.where(y[indices] == ( 1- minorityclasslabel))[0])/(K + 0.0)                                normConst = sum(rlist)                        for j in xrange(len(rlist)):                  normalizedrlist[j] = (rlist[j]/normConst)                        return normalizedrlist                    def get_indicesMinority(y, minorityclasslabel = 1):              y_new = []              for i in range(len(y)):                  if y[i] == 1:                      y_new.append(1)                  else:                      y_new.append(0)              y_new = np.asarray(y_new)              indicesMinority = np.where(y_new == minorityclasslabel)[0]                                return indicesMinority, y_new                    def generateSamples(X, y, minorityclasslabel = 1, K =5,beta = 0.3):              syntheticdata_X = []              syntheticdata_y = []                                          indicesMinority, y_new = get_indicesMinority(y)              ymin = y[indicesMinority]              Xmin = X[indicesMinority]                            rlist = getRis(X, y_new, indicesMinority, minorityclasslabel, K)              ml, ms = get_class_count(y_new)              G = getG(ml,ms, beta = beta)                           neigh = NearestNeighbors(n_neighbors=K)              neigh.fit(Xmin)                            for k in xrange(len(ymin)):                  g = int(np.round(rlist[k]*G))                            neighb_indx = neigh.kneighbors(Xmin[k],K,False)[0]                                        for l in xrange(g):                      ind = random.choice(neighb_indx)                      s = Xmin[k] + (Xmin[ind]-Xmin[k]) * random.random()                      syntheticdata_X.append(s)                      syntheticdata_y.append(ymin[k])                                    print 'asyn, raw X size:',X.shape                      X = np.vstack((X,np.asarray(syntheticdata_X)))                           y = np.hstack((y,syntheticdata_y))              print 'asyn, post X size:',X.shape                            return X , y
'''''
Created on 2014/03/09

@author: dylan
'''
from sklearn.neighbors import NearestNeighbors
import numpy as np
import random



def get_class_count(y, minorityclasslabel = 1):
    minorityclasslabel_count = len(np.where(y == minorityclasslabel)[0])
    maxclasslabel_count = len(np.where(y == (1 - minorityclasslabel))[0])

    return maxclasslabel_count, minorityclasslabel_count


# @param: X The datapoints e.g.: [f1, f2, ... ,fn]
# @param: y the classlabels e.g: [0,1,1,1,0,...,Cn]
# @param ms: The amount of samples in the minority group
# @param ml: The amount of samples in the majority group
# @return: the G value, which indicates how many samples should be generated in total, this can be tuned with beta
def getG(ml, ms, beta):
    return (ml-ms)*beta


# @param: X The datapoints e.g.: [f1, f2, ... ,fn]
# @param: y the classlabels e.g: [0,1,1,1,0,...,Cn]
# @param: minorityclass: The minority class
# @param: K: The amount of neighbours for Knn
# @return: rlist: List of r values
def getRis(X,y,indicesMinority,minorityclasslabel,K):

    ymin = np.array(y)[indicesMinority]
    Xmin = np.array(X)[indicesMinority]
    neigh = NearestNeighbors(n_neighbors= K)
    neigh.fit(X)

    rlist = [0]*len(ymin)
    normalizedrlist = [0]*len(ymin)

    for i in xrange(len(ymin)):
        indices = neigh.kneighbors(Xmin[i],K,False)[0]
#         print'y[indices] == (1 - minorityclasslabel):'
#         print y[indices]
#         print len(np.where(y[indices] == ( 1- minorityclasslabel))[0])
        rlist[i] = len(np.where(y[indices] == ( 1- minorityclasslabel))[0])/(K + 0.0)

    normConst = sum(rlist)

    for j in xrange(len(rlist)):
        normalizedrlist[j] = (rlist[j]/normConst)

    return normalizedrlist

def get_indicesMinority(y, minorityclasslabel = 1):
    y_new = []
    for i in range(len(y)):
        if y[i] == 1:
            y_new.append(1)
        else:
            y_new.append(0)
    y_new = np.asarray(y_new)
    indicesMinority = np.where(y_new == minorityclasslabel)[0]

    return indicesMinority, y_new

def generateSamples(X, y, minorityclasslabel = 1, K =5,beta = 0.3):
    syntheticdata_X = []
    syntheticdata_y = []


    indicesMinority, y_new = get_indicesMinority(y)
    ymin = y[indicesMinority]
    Xmin = X[indicesMinority]

    rlist = getRis(X, y_new, indicesMinority, minorityclasslabel, K)
    ml, ms = get_class_count(y_new)
    G = getG(ml,ms, beta = beta)

    neigh = NearestNeighbors(n_neighbors=K)
    neigh.fit(Xmin)

    for k in xrange(len(ymin)):
        g = int(np.round(rlist[k]*G))

        neighb_indx = neigh.kneighbors(Xmin[k],K,False)[0]

        for l in xrange(g):
            ind = random.choice(neighb_indx)
            s = Xmin[k] + (Xmin[ind]-Xmin[k]) * random.random()
            syntheticdata_X.append(s)
            syntheticdata_y.append(ymin[k])

    print 'asyn, raw X size:',X.shape
    X = np.vstack((X,np.asarray(syntheticdata_X)))

    y = np.hstack((y,syntheticdata_y))
    print 'asyn, post X size:',X.shape

    return X , y