Titanic: Machine Learning from Disaster

https://www.kaggle.com/c/titanic

2015.9.1
随机森林测试：
特征：
pclass+age+sibsp+parch+fare+sexnum+sex.name

树是50课时候，训练集90%正确率
15000课时候，测试集77.97%正确率，和给的例子结果一样

上面如果再增加其他原始特征，测试集的变差了。

特征的问题：

性别特征

年龄特征

家庭成员特征

组队特征

fare中有80%的数据都小于40

2015.9.9
方法：
决策树结果，随机森林结果，性别预测结果中生存两次的认为最终生存
结果：
准确率：0.8086124 预测对了338个 比之前的提高了7个
0 1
275 143

把两个年龄小于两岁的survive设为1，降低了，错了两个，这个所在的行号是：355和285

特征最重要，上面只是投票多的获胜。

setwd("E:/kaggle/titanic/data")train = read.csv("train.csv")test = read.csv("test.csv")test$survived <- 0comb<- rbind(train,test)comb$name <- as.character(comb$name)# strsplit(comb$name,split='[,.]')# strsplit(comb$name,split='[,.]')[1]# strsplit(comb$name,split='[,.]')[[1]][2]# 根据名字划分 名和姓 称号comb$surname <- sapply(comb$name,FUN=function(x){strsplit(x,split='[,.]')[[1]][1]})comb$title <-sapply(comb$name,FUN=function(x){strsplit(x,split='[,.]')[[1]][2]})comb$title <- sub(' ','',comb$title)comb$title[comb$title%in%c('Mme','Mlle')] <- 'Mlle'comb$title[comb$title%in%c('Capt','Don','Major','Sir')] <- 'Sir'comb$title[comb$title%in%c('Dona','Lady',"Countess",'Jonkheer')] <- 'Lady'comb$title <- as.factor(comb$title)comb$embarked[which(comb$embarked=='')] <- 'S'comb$embarked <- as.factor(comb$embarked)comb$fare[which(comb$fare=='')] <-median(comb$fare,na.rm=TRUE)comb$fare[which(comb$fare >= 40)] = 40 comb$family_size <- comb$sibsp + comb$parch + 1 comb$family_id <- paste(as.character(comb$family_size),comb$surname,sep="")comb$isfamily <- 2 comb$isfamily[which(comb$family_size==1)] = 1  library("rpart")library("rpart.plot")get_age_model <- rpart(age ~ pclass+sex+sibsp+parch+fare+embarked+title+family_size+isfamily,                        data = comb[!is.na(comb$age),],method="anova")comb$age[is.na(comb$age)] <- predict(get_age_model,comb[is.na(comb$age),])train_new <- comb[c(1:891),]test_new <- comb[c(892:1309),]# 本地测试# val = sample(1:nrow(train_new),0.6*nrow(train_new))# sub_train_new <- train_new[val,]# sub_test_new <- train_new[-val,]# tree_model <- rpart(survived~pclass+sex+sibsp+parch+fare+embarked+title+family_size+isfamily#                       #                     ,data = sub_train_new,method = "class",control=rpart.control(cp=0.0001))# # prp(tree_model,type=4,extra = 100)# # sub_test_new$predict <- predict(tree_model,sub_test_new,type = "class")# # acc = length(which(sub_test_new$survived == sub_test_new$predict))/nrow(sub_test_new)# # acc# 决策树tree_model <- rpart(survived~pclass+sex+sibsp+parch+fare+embarked+title+family_size                    ,data = train_new,method = "class",control=rpart.control(cp=0.0001))rpartsurvived <- predict(tree_model,test_new,type = "class")# result <- data.frame(PassengerId = test_new$passengerid, Survived = test_new$survived)# # write.csv(result,file="Submissions//rpart_add_family.csv",row.names=F)# 随机森林library(randomForest)train_new <- as.data.frame(train_new)train_new$survived <- as.factor(train_new$survived)model <- randomForest(survived~pclass+sex+sibsp+parch+fare+embarked+title+family_size,                      data=train_new,ntree=50,importance = TRUE)randomForestsurvived <- predict(model,test_new)gendersurvived <- read.csv("Submissions//genderclassmodel.csv")rpartsurvived <- as.data.frame(rpartsurvived)randomForestsurvived <- as.data.frame(randomForestsurvived)rpart_RF_gender <- cbind(rpartsurvived,randomForestsurvived,gendersurvived[,2])colnames(rpart_RF_gender) <- c("rpart","RF","gender")write.csv(rpart_RF_gender,file="Submissions//all_result_rpart_RF_gender.csv",row.names=F)# 按照行求和rpart_RF_gender<- read.csv("Submissions//all_result_rpart_RF_gender.csv")sum_class <- apply(rpart_RF_gender,1,sum)# 生存两次的为最后生存的val = which(sum_class>=2)summary_rpart_rf_gender<- data.frame(PassengerId = test_new$passengerid)summary_rpart_rf_gender$Survived <-0summary_rpart_rf_gender$Survived[val] <-1write.csv(summary_rpart_rf_gender,file="Submissions//summary_rpart_RF_gender.csv",row.names=F)# 结果 ：0.8086124 准确率 预测对了338个 比之前的提高了7个