機器學習與R之決策樹C50算法

決策樹
經(jīng)驗熵是針對所有樣本的分類結(jié)果而言
經(jīng)驗條件熵是針對每個特征里每個特征樣本分類結(jié)果之特征樣本比例和
基尼不純度
簡單地說就是從一個數(shù)據(jù)集中隨機選取子項，度量其被錯誤分類到其他分組里的概率

決策樹算法使用軸平行分割來表現(xiàn)具體一定的局限性
C5.0算法--可以處理數(shù)值型和缺失 只使用最重要的特征--使用的熵度量-可以自動修剪枝
劃分數(shù)據(jù)集
set.seed(123) #設(shè)置隨機種子
train_sample <- sample(1000, 900)#從1000里隨機900個數(shù)值
credit_train <- credit[train_sample, ]
credit_test  <- credit[-train_sample, ]
library(C50)
credit_model <- C5.0(credit_train[-17], credit_train$default) #特征數(shù)據(jù)框-標簽
C5.0(train,labers,trials = 1,costs = NULL)
trials控制自動法循環(huán)次數(shù)多迭代效果更好 costs可選矩陣 與各類型錯誤項對應(yīng)的成本-代價矩陣
summary(credit_model)#查看模型
credit_pred <- predict(credit_model, credit_test)#預(yù)測
predict(model,test,type="class")  type取class分類結(jié)果或者prob分類概率
單規(guī)則算法(1R算法)--單一規(guī)則直觀,但大數(shù)據(jù)底下,對噪聲預(yù)測不準
library(RWeka)
mushroom_1R <- OneR(type ~ ., data = mushrooms)
重復(fù)增量修建算法(RIPPER) 基于1R進一步提取規(guī)則
library(RWeka)

mushroom_JRip <- JRip(type ~ ., data = mushrooms)


[plain] view plain copy

    credit <- read.csv("credit.csv")  
    str(credit)  
      
    # look at two characteristics of the applicant  
    table(credit$checking_balance)  
    table(credit$savings_balance)  
      
    # look at two characteristics of the loan  
    summary(credit$months_loan_duration)  
    summary(credit$amount)  
      
    # look at the class variable  
    table(credit$default)  
      
    # create a random sample for training and test data  
    # use set.seed to use the same random number sequence as the tutorial  
    set.seed(123)  
    #從1000里隨機900個數(shù)值  
    train_sample <- sample(1000, 900)  
      
    str(train_sample)  
      
    # split the data frames切分數(shù)據(jù)集  
    credit_train <- credit[train_sample, ]  
    credit_test  <- credit[-train_sample, ]  
      
    # check the proportion of class variable類別的比例  
    prop.table(table(credit_train$default))  
    prop.table(table(credit_test$default))  
      
    ## Step 3: Training a model on the data ----  
    # build the simplest decision tree  
    library(C50)  
    credit_model <- C5.0(credit_train[-17], credit_train$default)  
      
    # display simple facts about the tree  
    credit_model  
      
    # display detailed information about the tree  
    summary(credit_model)  
      
    ## Step 4: Evaluating model performance ----  
    # create a factor vector of predictions on test data  
    credit_pred <- predict(credit_model, credit_test)  
      
    # cross tabulation of predicted versus actual classes  
    library(gmodels)  
    CrossTable(credit_test$default, credit_pred,  
               prop.chisq = FALSE, prop.c = FALSE, prop.r = FALSE,  
               dnn = c('actual default', 'predicted default'))  
      
    ## Step 5: Improving model performance ----  
      
    ## Boosting the accuracy of decision trees  
    # boosted decision tree with 10 trials提高模型性能 利用boosting提升  
    credit_boost10 <- C5.0(credit_train[-17], credit_train$default,  
                           trials = 10)  
    credit_boost10  
    summary(credit_boost10)  
      
    credit_boost_pred10 <- predict(credit_boost10, credit_test)  
    CrossTable(credit_test$default, credit_boost_pred10,  
               prop.chisq = FALSE, prop.c = FALSE, prop.r = FALSE,  
               dnn = c('actual default', 'predicted default'))  
      
    ## Making some mistakes more costly than others  
      
    # create dimensions for a cost matrix  
    matrix_dimensions <- list(c("no", "yes"), c("no", "yes"))  
    names(matrix_dimensions) <- c("predicted", "actual")  
    matrix_dimensions  
      
    # build the matrix設(shè)置代價矩陣  
    error_cost <- matrix(c(0, 1, 4, 0), nrow = 2, dimnames = matrix_dimensions)  
    error_cost  
      
    # apply the cost matrix to the tree  
    credit_cost <- C5.0(credit_train[-17], credit_train$default,  
                              costs = error_cost)  
    credit_cost_pred <- predict(credit_cost, credit_test)  
      
    CrossTable(credit_test$default, credit_cost_pred,  
               prop.chisq = FALSE, prop.c = FALSE, prop.r = FALSE,  
               dnn = c('actual default', 'predicted default'))  
      
    #### Part 2: Rule Learners -------------------  
      
    ## Example: Identifying Poisonous Mushrooms ----  
    ## Step 2: Exploring and preparing the data ---- 自動因子轉(zhuǎn)換--將字符標記為因子減少存儲  
    mushrooms <- read.csv("mushrooms.csv", stringsAsFactors = TRUE)  
      
    # examine the structure of the data frame  
    str(mushrooms)  
      
    # drop the veil_type feature  
    mushrooms$veil_type <- NULL  
      
    # examine the class distribution  
    table(mushrooms$type)  
      
    ## Step 3: Training a model on the data ----  
    library(RWeka)  
      
    # train OneR() on the data  
    mushroom_1R <- OneR(type ~ ., data = mushrooms)  
      
    ## Step 4: Evaluating model performance ----  
    mushroom_1R  
    summary(mushroom_1R)  
      
    ## Step 5: Improving model performance ----  
    mushroom_JRip <- JRip(type ~ ., data = mushrooms)  
    mushroom_JRip  
    summary(mushroom_JRip)  
      
    # Rule Learner Using C5.0 Decision Trees (not in text)  
    library(C50)  
    mushroom_c5rules <- C5.0(type ~ odor + gill_size, data = mushrooms, rules = TRUE)  
    summary(mushroom_c5rules)