PeterFankhauserIDS | a18d535 | 2021-04-18 10:38:27 +0200 | [diff] [blame] | 1 | library(caret) |
| 2 | library(tidyverse) |
| 3 | library(DMwR) |
| 4 | library(randomForest) |
| 5 | library(FSelector) |
| 6 | library(MLmetrics) |
| 7 | # library(randomForestExplainer) |
| 8 | # may need to: options(expressions = 5e5) to avoid stackoverflow for installing package |
| 9 | |
| 10 | set.seed(42) |
| 11 | |
| 12 | # Test |
| 13 | |
| 14 | ngramfile2<-"gold03_anno_ml_synfeat_nstopw" # 2nd dataset |
| 15 | ngramfile1 <-"goldstandard01_anno_ml_synfeat_nstop1" # 1st dataset |
| 16 | |
| 17 | setwd(dirname(rstudioapi::getSourceEditorContext()$path)) |
| 18 | stopwords <- readLines(con = "../data/stopwords.txt",encoding="UTF-8") |
| 19 | oringramme1 <- read.csv(paste("../data/",ngramfile1,".csv",sep=""), header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE) |
| 20 | oringramme2 <- read.csv(paste("../data/",ngramfile2,".csv",sep=""), header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE) |
| 21 | |
| 22 | |
| 23 | syfeaturenames <- read.csv("../data/syfeatures.tsv", header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE) |
| 24 | |
| 25 | featurenames <- read.csv("../data/features.tsv", header = TRUE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE) |
| 26 | |
| 27 | deleteStopwords = function(wl, stopwords = NULL) { |
| 28 | wl[!(wl %in% stopwords)] |
| 29 | } |
| 30 | |
| 31 | oringramme1 <- oringramme1 %>% |
| 32 | mutate(CO_IDIOM = ifelse(is.na(CO_IDIOM),0,CO_IDIOM)) %>% # treat NAs as 0 |
| 33 | filter(CO_IDIOM < 2) # just two classes: 0 no idiom, 1 idiom |
| 34 | |
| 35 | oringramme2 <- oringramme2 %>% |
| 36 | mutate(CO_IDIOM = ifelse(is.na(CO_IDIOM),0,CO_IDIOM)) %>% # treat NAs as 0 |
| 37 | filter(CO_IDIOM < 2) # just two classes: 0 no idiom, 1 idiom |
| 38 | |
| 39 | # Reduce number of classes, treat null values, add NSTOPW, change names for SY features and rename all features |
| 40 | # new featurenames. |
| 41 | |
| 42 | ngramme1 <- oringramme1 %>% |
| 43 | add_column(NSTOPW = sapply(oringramme1$tokens,function(x) length(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords)))) %>% |
| 44 | add_column(nstokens = sapply(oringramme1$tokens, function(x) paste(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords),collapse=" "))) %>% |
| 45 | # select(-matches("CO_TOKEN.*"), -tokens) %>% |
| 46 | add_column(CO_GRAM = sapply(oringramme1$tokens, function(x) length(unlist(strsplit(x," ")))))%>% |
| 47 | select(-matches("CO_TOKEN.*")) %>% # keep tokens for interpretability |
| 48 | mutate(across(matches(".rank.*"), ~ replace_na(.x, 1000))) %>% |
| 49 | mutate(across(c("dice", "lfmd", "llr", "ld", "pmi"), ~ replace_na(.x, min(.x) - 1))) %>% |
| 50 | rename_at(syfeaturenames$innames, ~ syfeaturenames[syfeaturenames$innames==.x,]$synames ) %>% |
| 51 | rename_at(featurenames$oldnames, ~ featurenames[featurenames$oldnames==.x,]$newnames ) %>% |
| 52 | mutate(across(everything(), ~ replace_na(.x, 0))) %>% |
| 53 | mutate(CO_IDIOM = as.factor(if_else(CO_IDIOM == 1, "idiom", "no_idiom"))) # just two classes: 0 no idiom, 1 idiom |
| 54 | |
| 55 | ngramme2 <- oringramme2 %>% |
| 56 | add_column(NSTOPW = sapply(oringramme2$tokens,function(x) length(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords)))) %>% |
| 57 | add_column(nstokens = sapply(oringramme2$tokens, function(x) paste(deleteStopwords(tolower(unlist(strsplit(x," "))),stopwords),collapse=" "))) %>% |
| 58 | # select(-matches("CO_TOKEN.*"), -tokens) %>% |
| 59 | select(-matches("CO_TOKEN.*")) %>% # keep tokens for interpretability |
| 60 | mutate(across(matches(".rank.*"), ~ replace_na(.x, 1000))) %>% |
| 61 | mutate(across(c("dice", "lfmd", "llr", "ld", "pmi"), ~ replace_na(.x, min(.x) - 1))) %>% |
| 62 | rename_at(syfeaturenames$innames, ~ syfeaturenames[syfeaturenames$innames==.x,]$synames ) %>% |
| 63 | rename_at(featurenames$oldnames, ~ featurenames[featurenames$oldnames==.x,]$newnames ) %>% |
| 64 | mutate(across(everything(), ~ replace_na(.x, 0))) %>% |
| 65 | mutate(CO_IDIOM = as.factor(if_else(CO_IDIOM == 1, "idiom", "no_idiom"))) # just two classes: 0 no idiom, 1 idiom |
| 66 | |
| 67 | # combine |
| 68 | |
| 69 | ngramme1<-subset(ngramme1,select=-c(CO_SONGS)) |
| 70 | ngramme <- rbind(ngramme1,ngramme2[colnames(ngramme1)]) |
| 71 | # ngramme<-ngramme1 |
| 72 | |
| 73 | ngramme<-ngramme1%>% distinct(nstokens,.keep_all=T) |
| 74 | ngramme1<-ngramme1%>% distinct(nstokens,.keep_all=T) |
| 75 | |
| 76 | # Optional |
| 77 | write.table(ngramme,file=paste("../data/","combined_noduplicates.tsv",sep=""), sep = "\t", quote=F) |
| 78 | |
| 79 | write.table(ngramme1,file=paste("../data/","dataset1_noduplicates.tsv",sep=""), sep = "\t", quote=F) |
| 80 | |
| 81 | |
| 82 | |
| 83 | # featuresets |
| 84 | |
| 85 | o_vars <- c("O_C2_N", "O_C2_SGT", "O_DEREKO", "O_GRAM", "O_NSTOPW") |
| 86 | o_vars_1 <- c("O_DEREKO", "O_GRAM", "O_NSTOPW") |
| 87 | o_vars_2 <- c("O_C2_N", "O_C2_SGT") |
| 88 | co_vars <- c("CO_VEC","CO_VEC_LEX") |
| 89 | sy_c1_vars <- c("SY_C1_C_L","SY_C1_C_R","SY_C1_DICE","SY_C1_LD","SY_C1_LDAF","SY_C1_LL","SY_C1_MI","SY_C1_MI_L","SY_C1_MI_R","SY_C1_MI2","SY_C1_MI3","SY_C1_NMI") |
| 90 | sy_c1_vars_1 <- c("SY_C1_LD","SY_C1_LDAF","SY_C1_LL","SY_C1_MI","SY_C1_MI3") |
| 91 | # sy_c1_vars_1 <- c("SY_C1_LD","SY_C1_LL","SY_C1_MI","SY_C1_MI3") |
| 92 | sy_c2_vars <- c("SY_C2_EXP","SY_C2_G","SY_C2_K","SY_C2_LD","SY_C2_LL","SY_C2_LMI","SY_C2_MI","SY_C2_MI3","SY_C2_T","SY_C2_Z") |
| 93 | sy_c2_vars_1 <- c("SY_C2_G","SY_C2_LD","SY_C2_LL","SY_C2_MI","SY_C2_MI3") |
| 94 | sy_w_vars <- c("SY_W_AVG","SY_W_CON","SY_W_MAX","SY_W_NSUM","SY_W_NSUM_AF") |
| 95 | # sy_w_vars_1 <- c("SY_W_AVG","SY_W_CON","SY_W_MAX","SY_W_NSUM") |
| 96 | sy_r_vars <- c("SY_C1_R","SY_W_R1","SY_W_R2","SY_R_D") |
| 97 | |
| 98 | all_vars <- c(o_vars,co_vars,sy_c1_vars,sy_c2_vars,sy_w_vars,sy_r_vars) |
| 99 | all_vars_1 <- c(o_vars, co_vars, sy_c1_vars_1, sy_c2_vars_1, sy_w_vars, sy_r_vars) |
| 100 | |
| 101 | |
| 102 | # formulae for training and testing rf |
| 103 | |
| 104 | all_fml <- as.formula(paste("CO_IDIOM ~ ", paste(all_vars, collapse= "+"))) |
| 105 | all_fml_1 <- as.formula(paste("CO_IDIOM ~ ", paste(all_vars_1, collapse= "+"))) |
| 106 | |
| 107 | sy_c1_fml_1 <- as.formula(paste("CO_IDIOM ~ ", paste(sy_c1_vars_1, collapse= "+"))) |
| 108 | sy_c2_fml_1 <- as.formula(paste("CO_IDIOM ~ ", paste(sy_c2_vars_1, collapse= "+"))) |
| 109 | sy_w_fml <- as.formula(paste("CO_IDIOM ~ ", paste(sy_c2_vars_1, collapse= "+"))) |
| 110 | sy_r_fml <- as.formula(paste("CO_IDIOM ~ ", paste(sy_r_vars, collapse= "+"))) |
| 111 | co_fml <- as.formula(paste("CO_IDIOM ~ ", paste(co_vars, collapse= "+"))) |
| 112 | o_fml_1 <- as.formula(paste("CO_IDIOM ~ ", paste(o_vars_1, collapse= "+"))) |
| 113 | o_fml <- as.formula(paste("CO_IDIOM ~ ", paste(o_vars, collapse= "+"))) |
| 114 | o_fml_2 <- as.formula(paste("CO_IDIOM ~ ", paste(o_vars_2, collapse= "+"))) |
| 115 | |
| 116 | # Simple train/test split |
| 117 | |
| 118 | set.seed(111) |
| 119 | trainRows <- sample(nrow(ngramme), nrow(ngramme)*0.8, replace = FALSE) |
| 120 | train <- ngramme[trainRows,] |
| 121 | test <- ngramme[setdiff(1:nrow(ngramme),trainRows),] |
| 122 | |
| 123 | cat("Random Forest\n") |
| 124 | |
| 125 | rf_classifier = randomForest(all_fml, train, importance=TRUE) |
| 126 | |
| 127 | # only SY features |
| 128 | # rf_classifier = randomForest(fmlasy, train, importance=TRUE) |
| 129 | |
| 130 | prediction_for_table <- predict(rf_classifier, test %>% select(-CO_IDIOM)) |
| 131 | |
| 132 | res <- confusionMatrix(prediction_for_table, test$CO_IDIOM, positive= "idiom",mode="everything") |
| 133 | print(res) |
| 134 | collected_results <- bind_cols("rf" = res$byClass) |
| 135 | |
| 136 | # Sensitivity is recall of class 1 |
| 137 | # Pos Pred Value is precision |
| 138 | varImpPlot(rf_classifier) |
| 139 | |
| 140 | cat("Random Forest with cutoff\n") |
| 141 | prediction_for_table <- predict(rf_classifier,test %>% select(-CO_IDIOM), cutoff = c(0.3, 0.7)) |
| 142 | res <- confusionMatrix(prediction_for_table,test$CO_IDIOM, positive = "idiom",mode="everything") |
| 143 | collected_results <- bind_cols(collected_results, "rf with cutoff" = res$byClass) |
| 144 | print(res) |
| 145 | |
| 146 | cat("With SMOTE resampled training data\n") |
| 147 | smoted.data <- SMOTE(all_fml, subset(train, select = c("CO_IDIOM", all_vars)), perc.over = 1200, perc.under = 100) |
| 148 | rf_classifier = randomForest(all_fml, smoted.data, importance=TRUE) |
| 149 | prediction_for_table <- predict(rf_classifier,test %>% select(-CO_IDIOM)) |
| 150 | res <- confusionMatrix(prediction_for_table,test$CO_IDIOM, positive = "idiom",mode="everything") |
| 151 | collected_results <- bind_cols(collected_results, "rf with SMOTE" = res$byClass) |
| 152 | print(res) |
| 153 | |
| 154 | cat("With SMOTE and cutoff\n") |
| 155 | prediction_for_table <- predict(rf_classifier,test %>% select(-CO_IDIOM), cutoff = c(0.3, 0.7)) |
| 156 | res <- confusionMatrix(prediction_for_table,test$CO_IDIOM, positive = "idiom",mode="everything") |
| 157 | collected_results <- bind_cols(collected_results, "rf with SMOTE and cutoff" = res$byClass) |
| 158 | print(res) |
| 159 | |
| 160 | collected_results <- collected_results %>% |
| 161 | round(3) %>% |
| 162 | add_column(measure = names(res$byClass)) %>% |
| 163 | column_to_rownames("measure") |
| 164 | |
| 165 | View(collected_results) |
| 166 | |
| 167 | # Analysing tradeoff between Fscore, Recall, Precision for various cutoffs |
| 168 | # full range from precision almost 100% to recall almost 100% |
| 169 | rf_classifier = randomForest(all_fml, train, importance=TRUE) |
| 170 | cvalues<-tibble() |
| 171 | for (c in c(seq(from=0.4,to=0.99,by=0.025),0.999)) { |
| 172 | prediction_for_table <- predict(rf_classifier, test %>% select(-CO_IDIOM), cutoff = c(1-c, c)) |
| 173 | conf<-confusionMatrix(prediction_for_table, test$CO_IDIOM, positive = "idiom") |
| 174 | cvalues <-bind_rows(cvalues, c(cutoff=c, conf$byClass)) |
| 175 | } |
| 176 | cvalues %>% |
| 177 | select(c("cutoff", "Recall", "Precision", "F1", "Specificity", "Balanced Accuracy")) %>% |
| 178 | pivot_longer(!cutoff, names_to=c("measure")) %>% |
| 179 | ggplot(aes(cutoff, value, colour=measure)) + geom_line(size=1) + |
| 180 | scale_x_continuous(breaks = scales::pretty_breaks(n = 10)) + |
| 181 | scale_y_continuous(breaks = scales::pretty_breaks(n = 10)) |
| 182 | |
| 183 | # cross validation |
| 184 | |
| 185 | featuresets<-data.frame(read.csv("../data/featuresets.tsv", header = FALSE, sep = "\t", dec=".", quote="", encoding="UTF-8",stringsAsFactors=FALSE)) |
| 186 | |
| 187 | collected_results<-tibble() |
| 188 | for (i in c(1:nrow(featuresets))) { |
| 189 | set.seed(325) |
| 190 | train_control <- trainControl(method="repeatedcv", number=5, repeats=5, verboseIter=T,classProbs = TRUE, |
| 191 | savePredictions = "final") |
| 192 | vars<-unlist(strsplit(featuresets[i,2],",")) |
| 193 | fml <- as.formula(paste("CO_IDIOM ~ ", paste(vars, collapse= "+"))) |
| 194 | rf <- train(fml, |
| 195 | data=ngramme1, |
| 196 | method='rf', |
| 197 | tuneGrid = data.frame(.mtry = floor(sqrt(length(vars)))), |
| 198 | cutoff=c(0.3,0.7), |
| 199 | trControl=train_control) |
| 200 | res<-confusionMatrix(rf$pred$pred, rf$pred$obs,mode="everything") |
| 201 | collected_results <-bind_rows(collected_results, c(features=featuresets[i,1], res$byClass)) |
| 202 | } |
| 203 | |
| 204 | collected_results1<-collected_results[c(2:16),c(1,6,7,8,12)] |
| 205 | collected_results1<-collected_results1%>% |
| 206 | mutate(across(everything(), ~ replace_na(.x, 0))) %>% |
| 207 | mutate(across(c(2:5), ~ round(as.numeric(.x),digits=3))) |
| 208 | |
| 209 | # Analysing tradeoff between Fscore, Recall, Precision for various cutoffs |
| 210 | # full range from precision almost 100% to recall almost 100% |
| 211 | |
| 212 | |
| 213 | vars<-unlist(strsplit(featuresets[2,2],",")) |
| 214 | fml <- as.formula(paste("CO_IDIOM ~ ", paste(vars, collapse= "+"))) |
| 215 | cvalues<-tibble() |
| 216 | for (c in c(seq(from=0.4,to=0.99,by=0.025),0.999)) { |
| 217 | set.seed(325) |
| 218 | train_control <- trainControl(method="repeatedcv", number=5, repeats=5, verboseIter=T,classProbs = TRUE, |
| 219 | savePredictions = "final") |
| 220 | rf <- train(fml, |
| 221 | data=ngramme1, |
| 222 | method='rf', |
| 223 | tuneGrid = data.frame(.mtry = floor(sqrt(length(vars)))), |
| 224 | cutoff=c(1-c,c), |
| 225 | trControl=train_control) |
| 226 | res<-confusionMatrix(rf$pred$pred, rf$pred$obs,mode="everything") |
| 227 | cvalues <-bind_rows(cvalues, c(cutoff=c, res$byClass)) |
| 228 | } |
| 229 | cvalues %>% |
| 230 | select(c("cutoff", "Recall", "Precision", "F1", "Specificity", "Balanced Accuracy")) %>% |
| 231 | pivot_longer(!cutoff, names_to=c("measure")) %>% |
| 232 | ggplot(aes(cutoff, value, colour=measure)) + geom_line(size=1) + |
| 233 | scale_x_continuous(breaks = scales::pretty_breaks(n = 10)) + |
| 234 | scale_y_continuous(breaks = scales::pretty_breaks(n = 10)) |
| 235 | |
| 236 | cvaluesall<-cvalues |
| 237 | |
| 238 | |
| 239 | # Using estimates by random forest on entire dataset |
| 240 | |
| 241 | library(randomForest) |
| 242 | rf_classifier_full = randomForest(all_fml_1, data=ngramme1, importance=TRUE, cutoff=c(0.3,0.7)) |
| 243 | rf_classifier_full |
| 244 | # class.error is 1 - recall |
| 245 | varImpPlot(rf_classifier_full) |
| 246 | |
| 247 | # Feature ranking |
| 248 | |
| 249 | # rf features as table |
| 250 | |
| 251 | # correlated features seem to split their rankings |
| 252 | |
| 253 | rfranks<-importance(rf_classifier_full)[,3:4] |
| 254 | |
| 255 | # ttest |
| 256 | |
| 257 | idioms<-ngramme1 %>% filter(CO_IDIOM == "idiom") |
| 258 | nonidioms<-ngramme1 %>% filter(CO_IDIOM != "idiom") |
| 259 | |
| 260 | m1<-mean(unlist(nonidioms["SY_C1_LD"]),na.rm=T) |
| 261 | |
| 262 | nonidioms1<-nonidioms%>%filter(SY_C1_LD > m1) |
| 263 | |
| 264 | mean(unlist(idioms["SY_C1_LD"]),na.rm=T) |
| 265 | |
| 266 | idioms1<-ngramme1 %>% filter(CO_IDIOM == "idiom") |
| 267 | idioms2<-ngramme2 %>% filter(CO_IDIOM == "idiom") |
| 268 | |
| 269 | ttestPvalues<-sapply(all_vars_1, |
| 270 | function(sel) t.test(idioms[sel],nonidioms[sel])$p.value) |
| 271 | |
| 272 | t.test(idioms["SY_C1_LL"],nonidioms["SY_C1_LL"]) |
| 273 | |
| 274 | ttestSignificance<-sapply(all_vars_1, |
| 275 | function(sel) { |
| 276 | p<-t.test(idioms[sel],nonidioms[sel])$p.value |
| 277 | if (p < 0.001) { |
| 278 | return("***") |
| 279 | } |
| 280 | if (p < 0.01) { |
| 281 | return("**") |
| 282 | } |
| 283 | if (p < 0.05) { |
| 284 | return ("*") |
| 285 | } |
| 286 | return(" ") |
| 287 | }) |
| 288 | |
| 289 | |
| 290 | ttestSignificance1<-sapply(all_vars_1, |
| 291 | function(sel) { |
| 292 | p<-t.test(idioms[sel],nonidioms1[sel])$p.value |
| 293 | if (p < 0.001) { |
| 294 | return("***") |
| 295 | } |
| 296 | if (p < 0.01) { |
| 297 | return("**") |
| 298 | } |
| 299 | if (p < 0.05) { |
| 300 | return ("*") |
| 301 | } |
| 302 | return(" ") |
| 303 | }) |
| 304 | |
| 305 | |
| 306 | # information gain |
| 307 | # multiply by 1000 to avoid undersized bins |
| 308 | # features are ranked individually no matter their correlation |
| 309 | igain<-information.gain(all_fml_1, data=ngramme1%>%mutate_at(all_vars_1, ~ . * 1000),unit="log2") |
| 310 | |
| 311 | # difference between means (positive or negative?) |
| 312 | |
| 313 | diffMeans<-sapply(all_vars_1,function(sel) mean(unlist(idioms[sel]),na.rm=T)-mean(unlist(nonidioms[sel]),na.rm=T)) |
| 314 | |
| 315 | diffMeansSign<-sapply(all_vars_1,function(sel) ifelse(mean(unlist(idioms[sel]),na.rm=T)-mean(unlist(nonidioms[sel]),na.rm=T)>0,"+","-")) |
| 316 | |
| 317 | diffMeansSign1<-sapply(all_vars_1,function(sel) ifelse(mean(unlist(idioms[sel]),na.rm=T)-mean(unlist(nonidioms1[sel]),na.rm=T)>0,"+","-")) |
| 318 | |
| 319 | |
| 320 | featurenames[,c("newnames","explanation")] |
| 321 | |
| 322 | featureRanks<-cbind(rownames(rfranks),rfranks,igain,ttestPvalues,ttestSignificance,diffMeans,diffMeansSign,ttestSignificance1,diffMeansSign1) |
| 323 | colnames(featureRanks)[1]<-"newnames" |
| 324 | |
| 325 | featureRanks<-merge(featureRanks,featurenames[,c("newnames","explanation")],by="newnames") |
| 326 | |
| 327 | |
| 328 | |
| 329 | #randomForestExplainer::explain_forest(rf_classifier ) |
| 330 | |
| 331 | # averate estimates and feature ranks over 10 runs |
| 332 | |
| 333 | errrate<-0 |
| 334 | conf<-matrix(0,2,3) |
| 335 | featureRanks<-matrix(0,4,length(vars)) |
| 336 | for (i in 1:10) { |
| 337 | rfc =randomForest(fmla, data=ngramme, importance=TRUE) |
| 338 | #rfc =randomForest(fmla, data=ngramme, importance=TRUE, cutoff=c(0.2, 0.8)) |
| 339 | errrate<-errrate+rfc$err.rate[100,1] |
| 340 | conf<-conf+rfc$confusion |
| 341 | featureRanks<-featureRanks+ |
| 342 | cbind(importance(rfc)[,3:4], |
| 343 | information.gain(fmla, data=ngramme%>%mutate_at(vars, ~ . * 1000),unit="log2"), |
| 344 | sapply(vars, |
| 345 | function(sel) t.test(idioms[sel],nonidioms[sel])$p.value)) |
| 346 | print(errrate/i) |
| 347 | conf1<-round( |
| 348 | rbind( |
| 349 | cbind(conf[,1:2]/i,(1-conf[,3]/i)*100), |
| 350 | c(100*diag(conf[,1:2])/colSums(conf[,1:2]),NA), |
| 351 | c(rowSums(conf[,1:2]/i),NA)),digits=2) |
| 352 | colnames(conf1)<-c("1","0","rec") |
| 353 | rownames(conf1)<-c("1","0","prec","sum") |
| 354 | print(conf1) |
| 355 | } |
| 356 | featureRanks<-featureRanks/10 |
| 357 | colnames(featureRanks)<-c("MeanDecreaseAccuracy","MeanDecreaseGini","InformationGain","Ttest") |
| 358 | |
| 359 | |
| 360 | |
| 361 | |