R/idiomclassification_mk_pf.R

library(caret)
library(tidyverse)
library(DMwR)
library(randomForest)
library(FSelector)
# library(randomForestExplainer)
# may need to: options(expressions = 5e5) to avoid stackoverflow for installing package

setwd(dirname(rstudioapi::getSourceEditorContext()$path))
stopwords <- readLines(con = "../data/stopwords.txt",encoding="UTF-8")
oringramme <- read.csv("../data/gold03_anno_ml_synfeat_nstopw.csv", header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)
syfeaturenames <- read.csv("../data/syfeatures.tsv", header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)
# syfeaturenames$navalue<-sapply(syfeaturenames$navalue,as.numeric)

deleteStopwords =  function(wl, stopwords = NULL) {
  wl[!(wl %in% stopwords)]
}

ngramme <- oringramme %>%
  add_column(NSTOPW = sapply(oringramme$tokens,function(x) length(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords)))) %>%
  # select(-matches("CO_TOKEN.*"), -tokens) %>%
  select(-matches("CO_TOKEN.*")) %>% # keep tokens for interpretability
  mutate(across(matches(".rank.*"), ~ replace_na(.x, 1000))) %>%
  mutate(across(c("dice", "lfmd", "llr", "ld", "pmi"), ~ replace_na(.x, min(.x) - 1))) %>%
  rename_at(syfeatures$innames, ~ syfeatures[syfeatures$innames==.x,]$synames ) %>%
  mutate(across(everything(), ~ replace_na(.x, 0))) %>%
  mutate(CO_IDIOM = as.factor(if_else(CO_IDIOM !=1, 0, 1))) # just two classes: 0 no idiom, 1 idiom

covars <- c("CO_LL", "CO_Z", "CO_G", "CO_T", "CO_LOGDICE", "CO_PMI", "CO_MI3", "CO_DEREKO", "CO_SGT", "CO_WIN5_VEC","CO_WIN5_VEC_AUTOSEM")
syvars <- c(syfeaturenames$synames,"NSTOPW")
vars <- c(covars,syvars)

fmla <- as.formula(paste("CO_IDIOM ~ ", paste(vars, collapse= "+")))
fmlaco <- as.formula(paste("CO_IDIOM ~ ", paste(covars, collapse= "+")))
fmlasy <- as.formula(paste("CO_IDIOM ~ ", paste(syvars, collapse= "+")))

# Simple train/test split

trainRows <- sample(nrow(ngramme), nrow(ngramme)*0.8, replace = FALSE)
train <- ngramme[trainRows,]
test <- ngramme[setdiff(1:nrow(ngramme),trainRows),]

rf_classifier = randomForest(fmla, train, ntree=100, mtry=10, importance=TRUE)

# only SY features
# rf_classifier = randomForest(fmlasy, train, ntree=100, mtry=10, importance=TRUE)

prediction_for_table <- predict(rf_classifier, test %>% select(-CO_IDIOM))

# different cutoff for prediction
# prediction_for_table <- predict(rf_classifier, test %>% select(-CO_IDIOM), cutoff = c(0.8, 0.2))

confusion <- table(observed=test$CO_IDIOM,predicted=prediction_for_table)
conf <- confusionMatrix(confusion)
print(conf)
varImpPlot(rf_classifier)

# optional resampling with smote

smoted.data <- SMOTE(fmla, subset(train, select = c("CO_IDIOM", vars)), perc.over = 1200, perc.under = 100)
rf_classifier = randomForest(fmla, smoted.data, ntree=100, mtry=4, importance=TRUE)
prediction_for_table <- predict(rf_classifier,test %>% select(-CO_IDIOM))
confusion <- table(observed=test$CO_IDIOM,predicted=prediction_for_table)
confusionMatrix(confusion)

# Using estimates by random forest on entire dataset

library(randomForest)
rf_classifier_full = randomForest(fmla, data=ngramme, ntree=100, mtry=2, importance=TRUE, cutoff=c(0.8,0.2))
rf_classifier_full
varImpPlot(rf_classifier_full)

# Feature ranking

# rf features as table

# correlated features seem to split their rankings

rfranks<-importance(rf_classifier_full)[,3:4]

# ttest

idioms<-ngramme %>% filter(CO_IDIOM==1)
nonidioms<-ngramme %>% filter(CO_IDIOM!=1)

ttestPvalues<-sapply(vars,
                     function(sel) t.test(idioms[sel],nonidioms[sel])$p.value)

# information gain
# multiply by 1000 to avoid undersized bins
# features are ranked individually not matter their correlation
igain<-information.gain(fmla, data=ngramme%>%mutate_at(vars, ~ . * 1000),unit="log2")

featureRanks<-cbind(rfranks,igain,ttestPvalues)

#randomForestExplainer::explain_forest(rf_classifier )

# averate estimates and feature ranks over 10 runs
errrate<-0
conf<-matrix(0,2,3)
featureRanks<-matrix(0,4,length(vars))
for (i in 1:10) {
  # rfc =randomForest(fmla, data=ngramme, ntree=100, importance=TRUE)
  rfc =randomForest(fmla, data=ngramme, ntree=100, importance=TRUE, cutoff=c(0.8,0.2))
  errrate<-errrate+rfc$err.rate[100,1]
  conf<-conf+rfc$confusion
  featureRanks<-featureRanks+
    cbind(importance(rfc)[,3:4],
          sapply(vars,
                 function(sel) t.test(idioms[sel],nonidioms[sel])$p.value),
          information.gain(fmla, data=ngramme%>%mutate_at(vars, ~ . * 1000),unit="log2"))
  print(errrate/i)
  conf1<-round(
    rbind(
      cbind(conf[,1:2]/i,(1-conf[,3]/i)*100),
      c(100*diag(conf[,1:2])/colSums(conf[,1:2]),NA),
      c(rowSums(conf[,1:2]/i),NA)),digits=2)
  colnames(conf1)<-c("0","1","rec")
  rownames(conf1)<-c("0","1","prec","sum")
  print(conf1)
}
featureRanks<-featureRanks/10