案例|基于贪心算法的特征选择

Posted 2021-04-25 R语言中文社区

greedy Algorithm Feature Selection

贪心算法（又称贪婪算法）是指，在对问题求解时，总是做出在当前看来是最好的选择。也就是说，不从整体最优上加以考虑， 它所做出的是在某种意义上的局部最优解。贪心算法不是对所有问题都能得到整体最优解，关键是贪心策略的选择，选择的贪心 策略必须具备无后效性，即某个状态以前的过程不会影响以后的状态，只与当前状态有关。

算法设计：

1. 初始化问题的目标值

2. while（实现优化目标的约束条件）{

   利用筛选策略，求出解空间的一个可行解

   }

3. 将所有可行解组合成目标解空间。

options(warn = -1)
require(magrittr)
require(dplyr)
require(glmnet)
# Greedy Algorithm
GreedyAlgorithm = function(dataSet) {
  # 基于逻辑回归,以AUC作为评价指标,采用贪心算法进行特征筛选
  # 
  # Args:
  #   dataSet: A dataframe that contains a feature "label"
  # 
  # Returns:
  #   A vector of selected features
  features = data.frame(name = colnames(dataSet)) %>%
    dplyr::filter(name != "label")  # select all features of the dataSet except "label"
  features = as.vector(features$name)    
  featureSelect = c("label")  # init the feature vector to be selected
  scoreBefore = data.frame()  # init the storage whice stores the (feature,aucScore) tuple from the end of each iteration 
  while((nrow(scoreBefore)<2||scoreBefore[length(scoreBefore),2]>
        scoreBefore[length(scoreBefore) - 1,2])&&nrow(scoreBefore)<length(features)){
    score = data.frame()
    for(feature in features){
      if(length(intersect(feature,featureSelect)) == 0){
        trainData = dataSet[,append(featureSelect,feature)]
        model = glm(label~.,family = "binomial",data = trainData,epsilon = 1e-10)
        prediction = predict(model,trainData)
        aucValue = auc(trainData$label,prediction)
        score = rbind(score,data.frame(feature = feature,aucValue = aucValue))
      }
    }
    featureSelect = unique(append(featureSelect,as.character(score[which.max(score$aucValue),1])))
    scoreBefore = rbind(scoreBefore,score[which.max(score$aucValue),])
  }
  featureSelect = head(featureSelect,length(featureSelect)-1)  # delete the last feature that can't fit the iteration condition 
  return(featureSelect[-1])  # reture the selected features except "label"
}

KS值表征了模型将正例和负例区分开来的能力。值越大，模型的预测准确性越好。通常情况下，KS>0.3即可认为模型有比较好的预测准确性。

KS值计算方法：

将所有样本根据预测得分从低到高排序均分成N组，分别计算这N组的实际好样本数、坏样本数、累积好样本数、累积坏样本数、累积好样本数占比、 累积坏样本数占比，差值。其中，实际好坏样本数分别为该组内的好坏样本数，累积好坏样本数为该组累积的好坏样本数，累积好坏样本数占比为 累积好坏样本数占总好坏样本数的比值，差值为累积坏样本数占比减去累计好样本数占比。KS指标为差值绝对值的最大值。

# ksValue
KsValue = function(prediction,n){
  # Compute the ks value of a model
  # 
  # Args:
  #   prediction: A vector that the prediction of a model
  #   n: The group number 
  # 
  # Returns:
  #   A vector that the difference value between the rate cumulative bad sample and the rate of cumulative good sample 
  dataResult = sort(prediction,decreasing = T)
  a = c()
  b = c()
  c = c()
  a[1] = 0
  b[1] = 0
  c[1] = 0
  if(length(dataResult)%%n==0){
    cut = length(dataResult)/n
    for (i in 2:(n+1)) {
      a[i] = sum(dataResult[(cut*(i-2)+1):(cut*(i-1))])
      b[i] = length(dataResult[(cut*(i-2)+1):(cut*(i-1))])-a[i]
    }
  }else{
    cut = round(length(dataResult)/n)
    for (i in 2:n) {
      a[i] = sum(dataResult[(cut*(i-2)+1):(cut*(i-1))])
      b[i] = length(dataResult[(cut*(i-2)+1):(cut*(i-1))])-a[i]
    }
    a[n+1] = sum(dataResult[(cut*(n-2)+1):(cut*(n-1))])
    b[n+1] = length(dataResult[(cut*(n-2)+1):(cut*(n-1))])-a[n+1]
  }
  c = abs(cumsum(a)/sum(a)-cumsum(b)/sum(b))
  return(c)
}

require(caret)
require(pROC)
data = read.csv("/data/workspace/Rworkspace/data_test.csv",encoding = "UTF-8")
data %<>%
  mutate(label = ifelse(target>30,1,0))
data = data[,-1]
data = data.frame(apply(data, 2, function(x) ifelse(is.na(x),median(x,na.rm = T),x)))
# 剔除近似常量的变量
# feature1 = nearZeroVar(data)
# data = data[,-feature1]
# 剔除相关度过高的自变量
# dataCor = cor(data)
# highcor = findCorrelation(dataCor,0.8)
# data = data[,-highcor]
# 利用贪心算法进行特征选择
# feature = GreedyAlgorithm(dataSet = data)
load("/data/workspace/Rworkspace/featureSelect.RData")  # 数据量较大,生成html过程中该算法比较耗时,所以直接加载测试时已经选取的特征
set.seed(521)
ind = base::sample(2,nrow(data),replace=T,prob=c(0.7,0.3))
trainData = data[ind==1,]
testData = data[ind==2,]
model = cv.glmnet(as.matrix(trainData[,feature]),trainData[,"label"],
                    family = "binomial",type.measure = "auc",alpha = 0,
                    lambda.min.ratio = 0.0001)
prediction = predict(model,as.matrix(testData[,feature]),s="lambda.min",type="response")
# compute ksValue
ksValue = KsValue(prediction,10)
par(mfrow = c(2,1))
plot(density(ksValue),type = 'l',main = "ksValue Plot",xlab = "cutPoint",ylab = "density_ks")
ks_value = max(ksValue)
text(.2,1.0,paste("ksValue = ",ks_value))
roc(testData$label, as.vector(prediction), auc = T,plot = T,print.auc=T)

## 
## Call:
## roc.default(response = testData$label, predictor = as.vector(prediction),     auc = T, plot = T, print.auc = T)
## 
## Data: as.vector(prediction) in 5130 controls (testData$label 0) < 429 cases (testData$label 1).
## Area under the curve: 0.7385

par(mfrow=c(1,1))