In sungcheolkim78/FiDEL: Fermi-Dirac based Ensemble Learning

knitr::opts_chunk$set(echo = TRUE)
library(data.table)
library(FDclassifieR)
library(dplyr)

Porto Seguro's Safe Driver Prediction

dirname <- 'data_porto_maxGini_0.29'
flist <- list.files(path=dirname, pattern='0.*.csv')
auclist <- (as.numeric(gsub(".csv", "", flist)) + 1)/2

# calculate prevalence from train data set
train <- fread(paste0(dirname, '/', 'train.csv'), na.strings=c("-1","-1.0"))
#train <- train %>%
#  mutate_at(vars(ends_with("cat")), as.factor) %>%
#  mutate_at(vars(ends_with("bin")), as.logical) %>%
#  mutate(target = as.factor(target))

rho <- sum(y == 1)/length(y)

# create prediction matrix
predictions <- matrix(nrow = 892816, ncol=length(flist))

i <- 1
for (f in flist) {
  tmp <- fread(paste0(dirname, '/', f))
  predictions[ ,i] <- tmp[[2]]
  i <- i + 1
}

colnames(predictions) <- mapply(paste0, rep('A',length(flist)), auclist)

dim(predictions)

head(predictions)

hist(predictions[, 7])

Test SUMMA

library(summa)
ksumma <- summa::summa(predictions, "rank")

print(max(ksumma@estimated_rank))
print(min(ksumma@estimated_rank))
hist(ksumma@estimated_rank)

rmax <- max(ksumma@estimated_rank)
rmin <- min(ksumma@estimated_rank)
submit <- data.table(id=tmp$id, target=(ksumma@estimated_rank - rmin)/(rmax-rmin))
fwrite(submit, file='submission.csv')

Test FD Ensemble

fde1 <- fdensemble(predictions)

fde1 <- predict_performance(fde1, auclist, rho, alpha=1)

print(max(fde1@estimated_rank))
print(min(fde1@estimated_rank))
hist(fde1@estimated_rank)

hist(fde1@predictions[ ,4])

submit <- data.table(id=tmp$id, target=fde1@estimated_prob)
fwrite(submit, file='submission.csv')

plot_cor(fde1, class_flag='positive')

cor(fde1@rank_matrix)

Build model on training data set

set.seed(1024)

# remove complete_rate < 0.9 columns
tmp <- train[, -c(23,26,28)]
tmp <- train[complete.cases(tmp), ]

# change categorical variables to one-hot vectors
cat_vars <- names(tmp)[grepl('_cat$', names(tmp))]
tmp <- tmp %>%
  sample_n(30000) %>%
  mutate_at(.vars = cat_vars, .funs = as.factor)
tmp <- model.matrix(~ . - 1, data = tmp)

tmp <- as.data.frame(tmp)
tmp$target <- as.factor(ifelse(tmp$target == 1, 'filed', 'not'))

inTraining0 <- createDataPartition(tmp$target, p = .75, list = FALSE)
training <- tmp[ inTraining0,]
testing  <- tmp[-inTraining0,]
testingY <- as_label(y[-inTraining0])

table(training$target)

model_list <- c('nnet', 'rda', 'svmLinear', 'svmRadial', 'pls', 'knn', 'earth', 'avNNet', 'mlp', 'nb', 'rf', 'rpart', 'ctree', 'C5.0', 'gbm', 'bayesglm', 'glm', 'glmnet', 'simpls')
t1 <- mtrainer(c('rda', 'pls', 'glm'), dataInfo = 'Porto')

t1 <- train(t1, target~., training, update=T)

t1 <- t1 %>%
  addmodel.mtrainer(c('rda', 'pls')) %>%
  train(target~., training)

plot(t1)

sungcheolkim78/FiDEL documentation built on Nov. 13, 2024, 7:58 a.m.