Machine Learning Algorithms in R

ml.R

# bagged Lasso and elastic-net regularized generalized linear models
library(foreach)
library(glmnet)

Tiqoo.bagging <- function(data.train, data.test, length_divisor=5, iterations=5000){
  
  predictions <- foreach(m=1:iterations,.combine=cbind) %do% {
    
    training_positions <- sample(nrow(data.train), size=floor((nrow(data.train)/length_divisor)))
    train_pos <- 1:nrow(data.train) %in% training_positions
    
    # train
    data.train.i <- data.train[train_pos,]
    X <- as.matrix(subset(data.train.i, select=-c(id,label)))
    y <- as.factor(data.train.i$label)
    fit <- glmnet(X, y, family="binomial", alpha=0, lambda=2^17)
    
    # predict
    Z <- data.matrix(subset(data.test,  select=-c(id,label)))
    z <- as.factor(data.test$label)
    predict(fit, s=2^17, Z, type="response")
  }
  
  rowMeans(predictions)
  
}

# CV Folds
library(cvTools)
library(randomForest)
library(gbm)
library(ada)
library(glmnet)
library(AUC)

# train model using N folds ...
Tiqoo.train <- function(data.train, method, withId=F, ...){
  numFolds=7
  folds <- cvFolds(nrow(data.train), numFolds)
  for (fold.i in 1:numFolds){
    cat("trainning fold: ", fold.i)
    indices.test  <- sort(folds$subset[folds$which==fold.i])
    indices.train <- sort(folds$subset[folds$which!=fold.i])
    if(withId){
      res <- method(data.train[indices.train,], data.train[indices.test,], fold.i, ...)
      print(res$pred)
    }else{
      data.fold.i.train <- subset(data.train[indices.train,], select=-c(id))
      data.fold.i.test  <- subset(data.train[indices.test,], select=-c(id))
      res <- method(data.fold.i.train, data.fold.i.test, ...)
      print(res$pred)
    }
  }
}

Tiqoo.glmnet <- function(data.train, data.test) {
  X <- as.matrix(subset(data.train, select=-c(label)))
  y <- as.factor(data.train$label)
  
  model.fit <- glmnet(X, y, family="binomial", alpha=0, lambda=2^17)

  Z <- data.matrix(subset(data.test,  select=-c(label)))
  z <- as.factor(data.test$label)
  
  pred.prob <- predict(model.fit, s=2^17, Z, type="response")
  pred.roc <- roc(pred.prob, z)
  list(model = model.fit, pred = auc(pred.roc))
}


Tiqoo.train.adaboost <- function(data.train, data.test, formula) {
  model.fit <- gbm(formula, data=data.train, dist="adaboost", n.trees=1000, cv.folds=5)
  best.iter <- gbm.perf(model.fit, method="cv")
  print(best.iter)
  pred.prob <- predict(model.fit, data.test, best.iter, type="link")
  pred.roc <- roc(pred.prob, as.factor(data.test$label))
  list(model = model.fit, pred = auc(pred.roc))
}

Tiqoo.train.randomForest <- function(data.train, data.test, formula) {
  model.fit <- randomForest(formula, data.train)
  pred.prob <- predict(model.fit, data.test, type="prob", ntree=1500)
  pred.roc <- roc(pred.prob[,2], as.factor(data.test$label))
  list(model = model.fit, pred = AUC::auc(pred.roc))
}

Tiqoo.train.ada <- function(data.train, data.test, formula) {
  model.fit <- ada(formula, data.train)
  pred.prob <- predict(model.fit, data.test, type="probs")
  pred.roc <- roc(pred.prob[,2], as.factor(data.test$label))
  list(model = model.fit, pred = auc(pred.roc))
}