code（回归）

library(stringi)
library(text2vec)
library(data.table)
library(magrittr)
library(purrr)
library(xgboost)
library(jiebaR)
library(readxl)
#读取数据

train <- read_xlsx("./data/train.xlsx")
test <- read_xlsx("./data/test.xlsx")
word_all <- rbind(train[,1:2], test)

#使用全切模式full，定义一个切词引擎
cutter <- worker(type = "full",bylines = T,stop_word = "d:/stop_words.utf8")
cutter$write = "NOFILE"
#对评论comment进行分词
all_words <- sapply(word_all$Discuss, function(x) cutter <= x)

# 定义预处理函数以及 tokenization 函数
prep_fun = tolower
tok_fun = word_tokenizer

it_token = itoken(all_words,
                  preprocessor = prep_fun,
                  tokenizer = tok_fun,
                  ids = word_all$Id,
                  progressbar = T)
vocab = create_vocabulary(it_token)

bigram_vectorizer = vocab_vectorizer(vocab)
dtm_token = create_dtm(it_token, bigram_vectorizer)

#加入tf-idf特征
# 定义模型
tfidf = TfIdf$new()
lsa = LSA$new(n_topics = 20)
dtm_token_tfidf_lsa = fit_transform(dtm_token, tfidf) %>%
  fit_transform(lsa)
#添加lsa特征
add.lsa <- function(m, lsa) {
  l <- fit_transform(m, lsa)
  colnames(l) <- ncol(l) %>% seq() %>% paste0("lsa.", .)
  cbind2(m, l)
}


###############################
###划分验证集和训练集
###############################
library(caret)
set.seed(2018)
sample_index <- createDataPartition(train$Score,p=0.85,list = F)
sample.train <- dtm_token[1:100000,][sample_index[,1],]
sample.val <- dtm_token[1:100000,][-sample_index[,1],]

dtrain <- sample.train %>%
  transform(tfidf) %>%
  add.lsa(lsa) %>%
  xgb.DMatrix(label = train$Score[sample_index])


dval <- sample.val %>%
  transform(tfidf) %>%
  add.lsa(lsa) %>%
  xgb.DMatrix(label = train$Score[-sample_index])


watchlist <- list(val = dval, train = dtrain)

# dtrain <- dtm_token[1:100000,] %>% 
#   transform(tfidf) %>% 
#   add.lsa(lsa) %>% 
#   xgb.DMatrix(label = train$Score)
param <- list(  objective           = "reg:linear", 
                booster             = "gbtree",
                eta                 = 0.1,
                max_depth           = 10, #8
                subsample           = 0.95,
                colsample_bytree    = 0.85,
                #num_class           = 7
                L1                  = 100
)


# param <- list(max_depth = 20, eta = 0.1, objective = "multi:softmax", eval_metric = "merror", nthread = 4, num_class = 6, verbose = 2)
# bst <- xgb.train(param, dtrain, nrounds = 500)
bst <- xgb.train( params              = param, 
                  data                = dtrain, 
                  nrounds             = 2000,
                  #early_stopping_rounds    = 200,
                  watchlist           = watchlist,
                  maximize            = TRUE,
                  eval_metric         = "rmse",
                  print_every_n       = 10
                  )




####################################
##########预测
####################################
dtest <- dtm_token[100001:130000,] %>% 
  transform(tfidf) %>% 
  add.lsa(lsa)

pre <- predict(bst,dtest,ntreelimit = 700)
result <- data.frame(ID = test$Id, Score = pre, stringsAsFactors = F)
write.csv(result,"result_0210.csv",row.names = F)