rm(list=ls(all=TRUE)) #start with empty workspace library(dplyr) library(tidyr) library(stringr) library(caret) library(caTools) #initialize my functions merge_sort <- function (eeg,events){ # merge events with data data <- merge(eeg,events,'id') data$value <- NULL #seperate id into frame number and id data <- separate(data,id,c('subject','series','frame'),sep='_') data$frame <- as.numeric(data$frame) # order based on id data <- data[order(data$frame),] #data$seriesno <- NULL data <- unite(data,col=id,subject,series,frame,sep='_',remove = TRUE) data } left_hem <- function(train, test){ # combine train and test data nrow_train <- nrow(train) nrow_test <- nrow(test) # create event columns in test data set so we can combine rows with training data test$HandStart <- 0 test$FirstDigitTouch <- 0 test$BothStartLoadPhase <- 0 test$LiftOff <- 0 test$Replace <- 0 test$BothReleased <- 0 # combine train and test data combi <- rbind(train,test) # split back into train and test data sets hem_train <- combi[1:nrow_train,] hem_test <- combi[(nrow_train+1):(nrow_train+nrow_test),] # remove event columns from training data hem_test$HandStart <- NULL hem_test$FirstDigitTouch <- NULL hem_test$BothStartLoadPhase <- NULL hem_test$LiftOff <- NULL hem_test$Replace <- NULL hem_test$BothReleased <- NULL # return two outputs as a list return(list(hem_train,hem_test)) } glm_regression <- function(train,test){ #formule for regression model formula1 <- HandStart ~ Fp1+F7+F3+Fz+FC5+FC1+T7+C3+Cz+TP9+CP5+CP1+P7+P3+Pz+PO9+O1+Oz+PO10+Fp2+F4+F8+FC2+FC6+C4+T8+CP2+CP6+TP10+P4+P8+O2 formula2 <- FirstDigitTouch ~ Fp1+F7+F3+Fz+FC5+FC1+T7+C3+Cz+TP9+CP5+CP1+P7+P3+Pz+PO9+O1+Oz+PO10+Fp2+F4+F8+FC2+FC6+C4+T8+CP2+CP6+TP10+P4+P8+O2 formula3 <- BothStartLoadPhase ~ Fp1+F7+F3+Fz+FC5+FC1+T7+C3+Cz+TP9+CP5+CP1+P7+P3+Pz+PO9+O1+Oz+PO10+Fp2+F4+F8+FC2+FC6+C4+T8+CP2+CP6+TP10+P4+P8+O2 formula4 <- LiftOff ~ Fp1+F7+F3+Fz+FC5+FC1+T7+C3+Cz+TP9+CP5+CP1+P7+P3+Pz+PO9+O1+Oz+PO10+Fp2+F4+F8+FC2+FC6+C4+T8+CP2+CP6+TP10+P4+P8+O2 formula5 <- Replace ~ Fp1+F7+F3+Fz+FC5+FC1+T7+C3+Cz+TP9+CP5+CP1+P7+P3+Pz+PO9+O1+Oz+PO10+Fp2+F4+F8+FC2+FC6+C4+T8+CP2+CP6+TP10+P4+P8+O2 formula6 <- BothReleased ~ Fp1+F7+F3+Fz+FC5+FC1+T7+C3+Cz+TP9+CP5+CP1+P7+P3+Pz+PO9+O1+Oz+PO10+Fp2+F4+F8+FC2+FC6+C4+T8+CP2+CP6+TP10+P4+P8+O2 # train the logistic regression model reg_model1 <- glm(formula1, data=train,family=binomial) reg_model2 <- glm(formula2, data=train,family=binomial) reg_model3 <- glm(formula3, data=train,family=binomial) reg_model4 <- glm(formula4, data=train,family=binomial) reg_model5 <- glm(formula5, data=train,family=binomial) reg_model6 <- glm(formula6, data=train,family=binomial) # create predictions my_predict1 <- predict(reg_model1,newdata=test,type='response') my_predict2 <- predict(reg_model2,newdata=test,type='response') my_predict3 <- predict(reg_model3,newdata=test,type='response') my_predict4 <- predict(reg_model4,newdata=test,type='response') my_predict5 <- predict(reg_model5,newdata=test,type='response') my_predict6 <- predict(reg_model6,newdata=test,type='response') # combine predictions to create a single dataframe my_solution <- data.frame(test$id,my_predict1,my_predict2,my_predict3, my_predict4,my_predict5,my_predict6) my_solution } set.seed(100) options(scipen=999) #no scientific numbers options (digits = 4) #hold decimal places to 4 # number of subjects you want to analyze total_subj <- 1 # sub-sample training data to reduce computational load (min 1, max 8) sub_sample <- 1 #initialize list list_subj_predictions <- list() # Loop through subjects for (j in 1:total_subj){ print(paste('Currently analysing subject',j)) list_subj_traindata <- list() # initialize list subject <- j # obtain all training data for (i in 1:sub_sample) { file_name_eeg <- paste('../input/train/subj',subject,'_series',i, '_data.csv',sep='') file_name_events <- paste('../input/train/subj',subject,'_series',i, '_events.csv',sep='') eeg <- read.csv(file_name_eeg, header=TRUE, colClasses=c("character",rep("numeric",32)))#read data.csv based on file name events <- read.csv(file_name_events,header=TRUE, colClasses=c("character",rep("numeric",6))) #read events.csv based on file name list_subj_traindata[[i]]<-merge_sort(eeg,events) rm(eeg,events) } #merge all training series data in 1 single data frame per subject train_variable_name <- paste('subj',subject,'_traindata',sep='') assign(train_variable_name,rbind_all(list_subj_traindata)) rm(list_subj_traindata) #obtain and merge test data from all series list_subj_testdata <- list() count <- 0 # i is series number for (i in 9:10) { count <- count + 1 file_name_eeg <- paste('../input/test/subj',subject,'_series',i, '_data.csv',sep='') eeg <- read.csv(file_name_eeg, header=TRUE, colClasses=c("character",rep("numeric",32))) list_subj_testdata[[count]]<- eeg rm(eeg) } #merge all series data in 1 single data frame per subject test_variable_name <- paste('subj',subject,'_testdata',sep='') assign(test_variable_name,rbind_all(list_subj_testdata)) rm(list_subj_testdata) # pre-processing of data # remove all EEG data from right hemisphere print('starting preprocessing') func_output <- left_hem(get(train_variable_name),get(test_variable_name)) assign(train_variable_name,as.data.frame(func_output[[1]])) assign(test_variable_name,as.data.frame(func_output[[2]])) rm(func_output) #Logistic regression regression print('now performing regression') list_subj_predictions[[j]]<- glm_regression(get(train_variable_name), get(test_variable_name)) rm(list=ls(pattern="^subj")) #remove variables in workspace to reduce memory usage } # combine all subject predictions into 1 dataframe for submission my_submission <- rbind_all(list_subj_predictions) my_submission[is.na(my_submission)] <- 0 names(my_submission) <- c('id','HandStart','FirstDigitTouch', 'BothStartLoadPhase','LiftOff','Replace','BothReleased') my_submission[,2:ncol(my_submission)]= round(my_submission[,2:ncol(my_submission)],digits=4) # reduce file size by limiting number of decimal places write.csv(my_submission,file='submission1.csv',row.names=FALSE)