extended with analysing both datasets.

Change-Id: I3a21aec66367846fd6a9cb7ff5f2c1ef9648d5b7
diff --git a/R/idiomclassification_wiki1.R b/R/idiomclassification_wiki1.R
new file mode 100644
index 0000000..4ff7719
--- /dev/null
+++ b/R/idiomclassification_wiki1.R
@@ -0,0 +1,276 @@
+library(caret)
+library(tidyverse)
+library(DMwR)
+library(randomForest)
+library(FSelector)
+# library(randomForestExplainer)
+# may need to: options(expressions = 5e5) to avoid stackoverflow for installing package
+
+set.seed(42)
+
+ngramfilegold1<-"goldstandard01_anno_ml_synfeat_nstop1"
+ngramfilegold2<-"gold03_anno_ml_synfeat_nstopw"
+ngramfiletest<-"wikilist_cleanup_syn"
+setwd(dirname(rstudioapi::getSourceEditorContext()$path))
+stopwords <- readLines(con = "../data/stopwords.txt",encoding="UTF-8")
+oringramme1 <- read.csv(paste("../data/",ngramfilegold1,".csv",sep=""), header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)
+oringramme2 <- read.csv(paste("../data/",ngramfilegold2,".csv",sep=""), header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)
+oringrammetest <- read.csv(paste("../data/",ngramfiletest,".csv",sep=""), header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)
+syfeaturenames <- read.csv("../data/syfeatures.tsv", header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)
+featurenames <- read.csv("../data/features.tsv", header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)
+
+deleteStopwords =  function(wl, stopwords = NULL) {
+  wl[!(wl %in% stopwords)]
+}
+
+oringramme1 <- oringramme1 %>%
+  mutate(CO_IDIOM = ifelse(is.na(CO_IDIOM),0,CO_IDIOM)) %>%  # treat NAs as 0
+  filter(CO_IDIOM < 2)   # just two classes: 0 no idiom, 1 idiom
+
+oringramme2 <- oringramme2 %>%
+  mutate(CO_IDIOM = ifelse(is.na(CO_IDIOM),0,CO_IDIOM)) %>%  # treat NAs as 0
+  filter(CO_IDIOM < 2)   # just two classes: 0 no idiom, 1 idiom
+
+# feature sets
+
+sy_w_vars <- c("SY_W_AVG","SY_W_CON","SY_W_MAX","SY_W_NSUM","SY_W_NSUM_AF")# sy_w_vars_1 <- c("SY_W_AVG","SY_W_CON","SY_W_MAX","SY_W_NSUM")
+sy_r_vars <- c("SY_C1_R","SY_W_R1","SY_W_R2","SY_R_D")
+sy_c1_vars_1 <- c("SY_C1_LD","SY_C1_LDAF","SY_C1_LL","SY_C1_MI","SY_C1_MI3")
+o_vars_1 <- c("O_NSTOPW","O_GRAM")
+sy_vars <- c(sy_c1_vars_1, sy_w_vars, sy_r_vars,o_vars_1)
+
+sy_featurenames<-subset(featurenames,newnames %in% sy_vars)
+
+# Reduce number of classes, treat null values, add NSTOPW, change names for SY features and rename all features
+# new featurenames.
+
+ngramme1 <- oringramme1 %>%
+  add_column(NSTOPW = sapply(oringramme1$tokens,function(x) length(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords)))) %>%
+  add_column(nstokens = sapply(oringramme1$tokens, function(x) paste(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords),collapse=" "))) %>%
+  # select(-matches("CO_TOKEN.*"), -tokens) %>%
+  add_column(CO_GRAM = sapply(oringramme1$tokens, function(x) length(unlist(strsplit(x," ")))))%>%
+  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(syfeaturenames$innames, ~ syfeaturenames[syfeaturenames$innames==.x,]$synames ) %>%
+  rename_at(featurenames$oldnames, ~ featurenames[featurenames$oldnames==.x,]$newnames ) %>%
+  mutate(across(everything(), ~ replace_na(.x, 0))) %>%
+  mutate(CO_IDIOM = as.factor(if_else(CO_IDIOM == 1, "idiom", "no_idiom"))) %>% # just two classes: 0 no idiom, 1 idiom
+  add_column(GOLD = 1) %>%
+  select(c("CO_IDIOM","tokens","nstokens","GOLD",all_of(sy_vars)))
+
+ngramme2 <- oringramme2 %>%
+  add_column(NSTOPW = sapply(oringramme2$tokens,function(x) length(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords)))) %>%
+  add_column(nstokens = sapply(oringramme2$tokens, function(x) paste(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords),collapse=" "))) %>%
+  # 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(syfeaturenames$innames, ~ syfeaturenames[syfeaturenames$innames==.x,]$synames ) %>%
+  rename_at(featurenames$oldnames, ~ featurenames[featurenames$oldnames==.x,]$newnames ) %>%
+  mutate(across(everything(), ~ replace_na(.x, 0))) %>%
+  mutate(CO_IDIOM = as.factor(if_else(CO_IDIOM == 1, "idiom", "no_idiom"))) %>% # just two classes: 0 no idiom, 1 idiom
+  add_column(GOLD = 1) %>%
+  select(c("CO_IDIOM","tokens","nstokens","GOLD",all_of(sy_vars)))
+
+# combine
+
+ngramme <- rbind(ngramme1,ngramme2[colnames(ngramme1)])
+# ngramme<-ngramme1
+# ngramme <- ngramme2[colnames(ngramme1)]
+ngramme<-ngramme%>% distinct(nstokens,.keep_all=T)
+
+ngramme1<-ngramme1%>% distinct(nstokens,.keep_all=T)
+
+
+# discard all ngrams with less than 2 non stopwords (no syntagmatic features possible)
+
+ngramme1 <- ngramme1 %>%
+   filter(O_NSTOPW > 1)
+
+# wiki ngrams
+
+wikingramme <- oringrammetest %>%
+  add_column(NSTOPW = sapply(oringrammetest$tokens,function(x) length(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords)))) %>%
+  add_column(nstokens = sapply(oringrammetest$tokens, function(x) paste(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords),collapse=" "))) %>%
+  add_column(CO_GRAM = sapply(oringrammetest$tokens, function(x) length(unlist(strsplit(x," ")))))%>%
+  select(-c(IDIOM,KERN)) %>% # keep tokens for interpretability
+  mutate(across(matches(".rank.*"), ~ replace_na(.x, 1000))) %>%
+  mutate(across(c("dice", "lfmd", "llr", "ld", "pmi"), ~ replace_na(.x, min(.x) - 1))) %>%
+  rename_at(syfeaturenames$innames, ~ syfeaturenames[syfeaturenames$innames==.x,]$synames ) %>%
+  rename_at(sy_featurenames$oldnames, ~ sy_featurenames[sy_featurenames$oldnames==.x,]$newnames ) %>%
+  mutate(across(everything(), ~ replace_na(.x, 0))) %>%
+  add_column(CO_IDIOM = as.factor("idiom")) %>%
+  add_column(GOLD = 0) %>%
+  select(c("CO_IDIOM","tokens","nstokens","GOLD",all_of(sy_vars)))
+
+# remove all with NSTOPW > 1 (no syntagmatic context available)
+
+wikingramme <- wikingramme %>%
+   filter(O_NSTOPW > 1)
+
+# remove duplicates after stopword exclusion
+
+wikingramme <- wikingramme %>% distinct(nstokens,.keep_all=T)
+
+# find duplicates by lower cased tokens without stopwords
+
+bothngramme <- merge(ngramme[,c("tokens","nstokens","CO_IDIOM")],wikingramme[,c("tokens","nstokens","CO_IDIOM")],by="nstokens")
+
+# 100% agreement ;)
+
+# combine
+
+allngramme <- rbind(ngramme1,wikingramme)
+
+
+# and again remove duplicates
+
+allngramme <- allngramme %>% distinct(nstokens,.keep_all=T)
+
+ngramme1 <- allngramme %>% filter(GOLD==1)
+wikingramme <- allngramme %>% filter(GOLD==0)
+
+write.table(allngramme,file=paste("../data/","dataset1_wiki_noduplicates.tsv",sep=""), sep = "\t", quote=F)
+
+
+# formulae for training and testing rf
+
+sy_fmla <- as.formula(paste("CO_IDIOM ~ ", paste(sy_vars, collapse= "+")))
+
+# Train/Test split
+
+# Training: 80% Gold Standard, no_idiom + 100% Gold Standard, idiom
+# Test: 20% Gold Standard, no_idiom + 100 % Wiki, idiom
+
+noidiomsgold <- ngramme1 %>% filter(CO_IDIOM=="no_idiom")
+idiomsgold <- ngramme1 %>% filter(CO_IDIOM=="idiom")
+
+set.seed(1120)
+trainRows <- sample(nrow(noidiomsgold), nrow(noidiomsgold)*0.8, replace = FALSE)
+train <- rbind(noidiomsgold[trainRows,],idiomsgold)
+test <- rbind(noidiomsgold[setdiff(1:nrow(noidiomsgold),trainRows),],wikingramme)
+
+cat("Random Forest\n")
+
+rf_classifier = randomForest(sy_fmla, train, importance=TRUE)
+
+prediction_for_table <- predict(rf_classifier, test %>% select(-CO_IDIOM))
+
+res <- confusionMatrix(prediction_for_table, test$CO_IDIOM, positive= "idiom",mode="everything")
+print(res)
+collected_results <- bind_cols("rf" = res$byClass)
+
+# Sensitivity is recall of class 1
+# Pos Pred Value is precision
+varImpPlot(rf_classifier)
+
+cat("Random Forest with cutoff\n")
+prediction_for_table <- predict(rf_classifier,test %>% select(-CO_IDIOM), cutoff = c(0.05, 0.95))
+res <- confusionMatrix(prediction_for_table,test$CO_IDIOM, positive = "idiom",mode="everything")
+collected_results <- bind_cols(collected_results, "rf with cutoff" = res$byClass)
+print(res)
+
+cat("With SMOTE resampled training data\n")
+smoted.data <- SMOTE(fmlasy, subset(train, select = c("CO_IDIOM", syvars)), perc.over = 1200, perc.under = 100)
+rf_classifier = randomForest(fmlasy, smoted.data, importance=TRUE)
+prediction_for_table <- predict(rf_classifier,test %>% select(-CO_IDIOM))
+res <- confusionMatrix(prediction_for_table,test$CO_IDIOM, positive = "idiom")
+collected_results <- bind_cols(collected_results, "rf with SMOTE" = res$byClass)
+print(res)
+
+cat("With SMOTE and cutoff\n")
+prediction_for_table <- predict(rf_classifier,test %>% select(-CO_IDIOM), cutoff = c(0.2, 0.8))
+res <- confusionMatrix(prediction_for_table,test$CO_IDIOM, positive = "idiom")
+collected_results <- bind_cols(collected_results, "rf with SMOTE and cutoff" = res$byClass)
+print(res)
+
+collected_results <- collected_results %>%
+  round(3) %>%
+  add_column(measure = names(res$byClass)) %>%
+  column_to_rownames("measure")
+
+View(collected_results)
+
+
+# Analysing tradeoff between Fscore, Recall, Precision for various cutoffs
+# full range from precision almost 100% to recall almost 100%
+rf_classifier = randomForest(sy_fmla, train, importance=TRUE)
+cvaluesw<-tibble()
+for (c in c(seq(from=0.4,to=0.99,by=0.025),0.999)) {
+  prediction_for_table <- predict(rf_classifier, test %>% select(-CO_IDIOM), cutoff = c(1-c, c))
+  conf<-confusionMatrix(prediction_for_table, test$CO_IDIOM, positive = "idiom")
+  cvaluesw <-bind_rows(cvaluesw, c(cutoff=c, conf$byClass))
+}
+cvaluesw %>%
+  select(c("cutoff", "Recall", "Precision", "F1", "Specificity", "Balanced Accuracy")) %>%
+  pivot_longer(!cutoff, names_to=c("measure")) %>%
+  ggplot(aes(cutoff, value, colour=measure)) + geom_line(size=1) +
+  scale_x_continuous(breaks = scales::pretty_breaks(n = 10)) +
+  scale_y_continuous(breaks = scales::pretty_breaks(n = 10))
+
+
+# Using estimates by random forest on entire dataset
+
+library(randomForest)
+rf_classifier_full = randomForest(fmlasy, data=allngramme, importance=TRUE)
+rf_classifier_full
+# class.error is 1 - recall
+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 == "idiom")
+nonidioms<-ngramme %>% filter(CO_IDIOM != "idiom")
+
+ttestPvalues<-sapply(syvars,
+                     function(sel) t.test(idioms[sel],nonidioms[sel])$p.value)
+
+# information gain
+# multiply by 1000 to avoid undersized bins
+# features are ranked individually no matter their correlation
+igain<-information.gain(fmlasy, data=ngramme%>%mutate_at(syvars, ~ . * 1000),unit="log2")
+
+featureRanks<-cbind(rfranks,igain,ttestPvalues)
+
+#randomForestExplainer::explain_forest(rf_classifier )
+
+# averate estimates and feature ranks over 10 runs
+
+errrate<-0
+conf<-matrix(0,2,3)
+featureRanks<-matrix(0,4,length(vars))
+for (i in 1:10) {
+  rfc =randomForest(fmlasy, data=ngramme, importance=TRUE)
+  #rfc =randomForest(fmlasy, data=ngramme, importance=TRUE, cutoff=c(0.2, 0.8))
+  errrate<-errrate+rfc$err.rate[100,1]
+  conf<-conf+rfc$confusion
+  featureRanks<-featureRanks+
+    cbind(importance(rfc)[,3:4],
+          information.gain(fmlasy, data=ngramme%>%mutate_at(syvars, ~ . * 1000),unit="log2"),
+          sapply(syvars,
+                 function(sel) t.test(idioms[sel],nonidioms[sel])$p.value))
+  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("1","0","rec")
+  rownames(conf1)<-c("1","0","prec","sum")
+  print(conf1)
+}
+featureRanks<-featureRanks/10
+colnames(featureRanks)<-c("MeanDecreaseAccuracy","MeanDecreaseGini","InformationGain","Ttest")
+
+
+
+