comparison evaluate_aggregation.R @ 91:b0b172279433 draft

Uploaded

91:b0b172279433

########################################################
#
# creation date : 10/10/16
# last modification : 29/10/16
# author : Dr Nicolas Beaume
#
########################################################

suppressWarnings(suppressMessages(library(GA)))
library("miscTools")
library(rpart)
suppressWarnings(suppressMessages(library(randomForest)))
library(e1071)
suppressWarnings(suppressMessages(library(glmnet)))
library(rrBLUP)
options(warn=-1)
############################ helper functions #######################

##### classifiers
prediction <- function(genotype, model, classifier="unknown") {
  # run prediction according to the classifier
  switch(classifier,
         rrBLUP={
           predictions <- as.matrix(genotype) %*% as.matrix(model$u);
           predictions <- predictions[,1]+model$beta;
         },
         rf={
           predictions <- predict(model, genotype);
         },
         svm={
           predictions <- predict(model, genotype);
         },
         lasso={
           predictions <- predict(model, as.matrix(genotype), type = "response");
         },
         {warning("unkonwn classifier, please choose among the following : rrBLUP, rf, svm, lasso")})
  return(predictions)
}

# extract parameter from a model, excluding rrBLUP which auto-optimize
extractParameter <- function(model, classifierName) {
  param <- NULL
  switch(classifierName,
         # random forest
         rf={
          param <- model$ntree
          param <- c(param, list(model$mtry))
          names(param) <- c("ntree", "mtry")
           },
         # svm
         svm={
          param <- as.numeric(model$cost)
          param <- c(param, list(model$gamma))
          param <- c(param, list(model$coef0))
          param <- c(param, list(model$degree))
          param <- c(param, list(model$kernel))
          names(param) <- c("c", "g", "coef", "d", "kernel")
          switch((model$kernel+1),
                 param$kernel <- "linear",
                 param$kernel <- "polynomial",
                 param$kernel <- "radial",
                 param$kernel <- "sigmoid"
                 )
           },
         # lasso
         lasso={
           param <- as.list(model$lambda)
           names(param) <- "lambda"
          },
         {print(paste("unknown classifier, please choose among rf, svm, lasso"));
           stop()}
  )
  return(param)
}

##### Genetic algorithm

# compute r2 by computing the classic formula
# compare the sum of square difference from target to prediciton
# to the sum of square difference from target to the mean of the target
r2 <- function(target, prediction) {
  sst <- sum((target-mean(target))^2)
  ssr <- sum((target-prediction)^2)
  return(1-ssr/sst)
}

optimizeOneIndividual <- function(values, trueValue) {
  # change the value into a function
  f <- function(w) {sum(values * w/sum(w))}
  fitness <- function(x) {1/abs(trueValue-f(x))}
  resp <- ga(type = "real-valued", fitness = fitness, min = rep(0, length(values)), max = rep(1, length(values)), 
             maxiter = 1000, monitor = NULL, keepBest = T)
  resp@solution <- resp@solution/sum(resp@solution)
  return(resp)
}

optimizeWeight <- function(values, trueValue, n=1000) {
  fitnessAll <- function(w) {
    predicted <- apply(values, 1, weightedPrediction.vec, w)
    return(mean(r2(trueValue, predicted)))
    #return(mean(1/abs(trueValue-predicted)))
  }
  resp <- ga(type = "real-valued", fitness = fitnessAll, min = rep(0, ncol(values)), max = rep(1, ncol(values)), 
             maxiter = n, monitor = NULL, keepBest = T)
  resp@solution <- resp@solution/sum(resp@solution)
  return(resp)
}

weightedPrediction <- function(classifiers, w) {
  if(length(w) > ncol(classifiers)) {
    warning("more weights than classifiers, extra weigths are ignored")
    w <- w[1:ncol(classifiers)]
  } else if(length(w) < ncol(classifiers)) {
    warning("less weights than classifiers, extra classifiers are ignored")
    classifiers <- classifiers[,1:length(w)]
  }
  prediction <- NULL
  prediction <- c(prediction, apply(classifiers, 1, weightedPrediction.vec, w))
  return(prediction)
}

weightedPrediction.vec <- function(values, w) {
  return(sum(values * w/sum(w)))
}

##### meta-decision tree

tuneTree <- function(data, target) {
  data <- data.frame(data, target=target)
  size <-  nrow(data)
  xerror <-  NULL
  split <-  1:ceiling(size/5)
  leafSize <-  1:ceiling(size/10)
  xerror <- matrix(rep(-1, length(split)*length(leafSize)), ncol=length(leafSize))
  cp <- matrix(rep(-1, length(split)*length(leafSize)), ncol=length(leafSize))
  for(i in 1:length(split)) {
    for(j in 1:length(leafSize)) {
      op <- list(minsplit=split[i], minbucket=leafSize[j])
      tree <- rpart(target ~., data=data, control=op, method="anova")
      xerror[i,j] <- tree$cptable[which.min(tree$cptable[,"xerror"]),"xerror"]
      cp[i,j] <- tree$cptable[which.min(tree$cptable[,"xerror"]),"CP"]
    }
  }
  index <- which(xerror==min(xerror), arr.ind = T)
  op <- list(minsplit=split[index[1]], minbucket=leafSize[index[2]], cp=cp[index[1], index[2]])
  return(op)
}

###### meta-LASSO
# create fold by picking at random row indexes
createFolds <- function(nbObs, n) {
  # pick indexes
  index <- sample(1:n, size=nbObs, replace = T)
  # populate folds
  folds <- NULL
  for(i in 1:n) {
    folds <- c(folds, list(which(index==i)))
  }
  return(folds)
}

searchParamLASSO <- function(genotype, phenotype, alpha=seq(0,1,0.1), n=7) {
  folds <- createFolds(nrow(genotype), n = n)
  acc <- NULL
  indexAlpha <- 1
  for(a in alpha) {
    curAcc <- NULL
    for(i in 1:n) {
      train <- genotype[-folds[[i]],]
      test <- genotype[folds[[i]],]
      phenoTrain <- phenotype[-folds[[i]]]
      phenoTest <- phenotype[folds[[i]]]
      cv <- cv.glmnet(x=as.matrix(train), y=phenoTrain, alpha=a)
      model <- glmnet(x=as.matrix(train), y=phenoTrain, alpha=a, lambda = cv$lambda.1se)
      pred <- predict(model, test, type = "response")
      curAcc <- c(curAcc, r2(phenoTest, pred))
    }
    acc <- c(acc, mean(curAcc))
  }
  names(acc) <- alpha
  return(as.numeric(names(acc)[which.max(acc)]))
}

###### meta-random forest

searchParamRF <- function(genotype, phenotype, rangeNtree, mtry=ncol(genotype)) {
  n <- ceiling(nrow(genotype)/3)
  indexTest <- sample(1:nrow(genotype), size=n)
  train <- genotype[-indexTest,]
  test <- genotype[indexTest,]
  phenoTrain <- phenotype[-indexTest]
  phenoTest <- phenotype[indexTest]
  acc <- NULL
  indexNtree <- 1
  for(ntree in rangeNtree) {
    model <- randomForest(x=train, y=phenoTrain, ntree = ntree, mtry = mtry)
    pred <- predict(model, test)
    acc <- c(acc, r2(phenoTest, pred))
  }
  names(acc) <- rangeNtree
  best <- which.max(acc)
  return(as.numeric(names(acc)[best]))
}

###### meta-SVM
searchParamSVM <- function(train, target, kernel="radial") {
  # tuning parameters then train
  model <- NULL
  switch(kernel,
         sigmoid={
           tune <-  tune.svm(train, target, gamma = 10^(-6:-1), cost = 10^(0:2), kernel="sigmoid");
           g <- tune$best.parameters[[1]];
           c <- tune$best.parameters[[2]];
           model <-  svm(x=train, y=target, gamma = g, cost = c, kernel = "sigmoid")},
         linear={
           tune <-  tune.svm(train, target, cost = 10^(0:2), kernel="linear");
           c <- tune$best.parameters[[1]];
           model <-  svm(x=train, y=target, cost = c, kernel = "linear")},
         polynomial={
           tune <-  tune.svm(train, target, gamma = 10^(-6:-1), cost = 10^(0:2), degree = 0:4, coef0 = 0:3, kernel="polynomial");
           d <- tune$best.parameters[[1]];
           g <- tune$best.parameters[[2]];
           coef <- tune$best.parameters[[3]];
           c <- tune$best.parameters[[4]];
           model <-  svm(x=train, y=target, gamma = g, cost = c, kernel = "polynomial", degree = d, coef0 = coef)},
         {
           tune <-  tune.svm(train, target, gamma = 10^(-6:-1), cost = 10^(0:3), kernel="radial");
           g <- tune$best.parameters[[1]];
           c <- tune$best.parameters[[2]];
           model <-  svm(x=train, y=target, gamma = g, cost = c, kernel = "radial")}
  )
  return(model)
}

#################### upper level functions #####################

aggregateDT <- function(train, test, target, folds) {
  r2Aggreg <- NULL
  for (i in 1:length(folds)) {
    trainIndex <- unlist(folds[-i])
    testIndex <- folds[[i]]
    treeParam <- tuneTree(train[[i]], target[trainIndex])
    data <- data.frame(train[[i]], target=target[trainIndex])
    model <- rpart(target ~., data=data, method = "anova", control = treeParam)
    model <- prune(model, cp=treeParam["cp"])
    pred <- predict(model, data.frame(test[[i]]))
    r2Aggreg <- c(r2Aggreg, r2(target[testIndex], pred))
  }
  return(r2Aggreg)
}

aggregateGeneticMean <- function(train, test, target, folds) {
  r2Aggreg <- NULL
  for (i in 1:length(folds)) {
    trainIndex <- unlist(folds[-i])
    testIndex <- folds[[i]]
    opt <- optimizeWeight(values = train[[i]], trueValue = target[trainIndex])
    pred <- weightedPrediction(test[[i]], opt@solution)
    r2Aggreg <- c(r2Aggreg, r2(target[testIndex], pred))
  }
  return(r2Aggreg)
}

aggregateLASSO <- function(train, test, target, folds) {
  r2Aggreg <- NULL
  for (i in 1:length(folds)) {
    trainIndex <- unlist(folds[-i])
    testIndex <- folds[[i]]
    alpha <- searchParamLASSO(train[[i]], target[trainIndex])
    cv <- cv.glmnet(x=as.matrix(train[[i]]), y=target[trainIndex], alpha=alpha)
    model <- glmnet(x=as.matrix(train[[i]]), y=target[trainIndex], alpha=alpha, lambda = cv$lambda.1se)
    pred <- predict(model, test[[i]])
    r2Aggreg <- c(r2Aggreg, r2(target[testIndex], pred))
  }
  return(r2Aggreg)
}

aggregateRF <- function(train, test, target, folds) {
  r2Aggreg <- NULL
  for (i in 1:length(folds)) {
    trainIndex <- unlist(folds[-i])
    testIndex <- folds[[i]]
    ntree <- searchParamRF(genotype = train[[i]], phenotype = target[trainIndex],
                           rangeNtree = seq(100, 1000, 100))
    model <- randomForest(x=as.matrix(train[[i]]), y=target[trainIndex],
                          ntree = ntree, mtry = ncol(train[[i]]))
    pred <- predict(model, test[[i]])
    r2Aggreg <- c(r2Aggreg, r2(target[testIndex], pred))
  }
  return(r2Aggreg)
}

aggregateSVM <- function(train, test, target, folds, kernel="linear") {
  r2Aggreg <- NULL
  for (i in 1:length(folds)) {
    trainIndex <- unlist(folds[-i])
    testIndex <- folds[[i]]
    model <- searchParamSVM(train = train[[i]], target = target[trainIndex], kernel = kernel)
    pred <- predict(model, test[[i]])
    r2Aggreg <- c(r2Aggreg, r2(target[testIndex], pred))
  }
  return(r2Aggreg)
}

################################### main #############################
# # load argument
cmd <- commandArgs(T)
source(cmd[1])
# load folds
con = file(folds)
folds <- readLines(con = con, n = 1, ok=T)
close(con)
folds <- readRDS(folds)
# phenotype is written as a table (in columns) but it must be sent as a vector for mixed.solve
phenotype <- read.table(phenotype, sep="\t", h=T)[,1]
# load genotype
con = file(genotype)
genotype <- readLines(con = con, n = 1, ok=T)
close(con)
genotype <- read.table(genotype, sep="\t", h=T)
# find which classifiers will be used for aggregation
classifNames <- NULL
if(lassoModel !="None"){
   classifNames <- c(classifNames, "lasso")
   con = file(lassoModel)
   lassoModel <- readLines(con = con, n = 1, ok=T)
   close(con)
   lassoModel <- readRDS(lassoModel)
}
if(rrBLUPModel !="None"){
  classifNames <- c(classifNames, "rrBLUP")
  con = file(rrBLUPModel)
  rrBLUPModel <- readLines(con = con, n = 1, ok=T)
  close(con)
  rrBLUPModel <- readRDS(rrBLUPModel)
}
if(rfModel !="None"){
  classifNames <- c(classifNames, "rf")
  con = file(rfModel)
  rfModel <- readLines(con = con, n = 1, ok=T)
  close(con)
  rfModel <- readRDS(rfModel)
}
if(svmModel !="None"){
  classifNames <- c(classifNames, "svm")
  con = file(svmModel)
  svmModel <- readLines(con = con, n = 1, ok=T)
  close(con)
  svmModel <- readRDS(svmModel)
}

# compute prediction of the training set and test set for each fold and each classifiers
# train predictions and test prediction are stored in separate lists 
# where each element of the list represent a folds
predictionTrain.list <- NULL
predictionTest.list <- NULL
r2Classif.list <- NULL 
for(i in 1:length(folds)) {
  # for the current fold, create training set and test set
  trainIndex <- unlist(folds[-i])
  testIndex <- folds[[i]]
  train <- genotype[trainIndex,]
  phenoTrain <- phenotype[trainIndex]
  test <- genotype[testIndex,]
  phenoTest <- phenotype[testIndex]
  # only to intialize data frame containing predictions
  predictionTrain <- matrix(rep(-1, length(classifNames)*length(trainIndex)), 
                            ncol=length(classifNames))
  colnames(predictionTrain) <- classifNames
  predictionTest <- matrix(rep(-1, length(classifNames)*length(testIndex)), 
                           ncol=length(classifNames))
  colnames(predictionTest) <- classifNames
  r2Classif <- NULL
  # for each classifiers, compute prediction on both sets 
  # and evaluate r2 to find the best classifier
  for(j in 1:length(classifNames)) {
    switch(classifNames[j],
           # random forest
           rf={
            # predict train and test
            param <- extractParameter(rfModel, "rf")
            model <- randomForest(x=train, y=phenoTrain, mtry = param$mtry, 
                                  ntree = param$ntree); 
            predictionTrain[,j] <- prediction(train, model, classifier = "rf");
            predictionTest[,j] <- prediction(test, model, classifier = "rf");
            r2Classif <- c(r2Classif, rf=r2(phenoTest, predictionTest[,"rf"]))},
           # svm
           svm={
            # predict train and test
            param <- extractParameter(svmModel, "svm");
            model <- svm(train, phenoTrain, kernel = param$kernel, cost = param$c,
                         gamma=param$g, degree = param$d, coef0 = param$coef, scale = F)
            predictionTrain[,j] <- prediction(train, model, classifier = "svm");
            predictionTest[,j] <- prediction(test, model, classifier = "svm");
            r2Classif <- c(r2Classif, svm=r2(phenoTest, predictionTest[,"svm"]))},
           # lasso
           lasso={
            # predict train and test
            param <- extractParameter(lassoModel, "lasso");
            model <- glmnet(x=as.matrix(train), y=phenoTrain, lambda = param$lambda);
            predictionTrain[,j] <- prediction(train, model, classifier = "lasso");
            predictionTest[,j] <- prediction(test, model, classifier = "lasso");
            r2Classif <- c(r2Classif, lasso=r2(phenoTest, predictionTest[,"lasso"]))},
           # rrBLUP
           rrBLUP={
            # predict train and test
            model <- mixed.solve(phenoTrain, Z=train,K=NULL, SE=F,return.Hinv = F);
            predictionTrain[,j] <- prediction(train, model, classifier = "rrBLUP");
            predictionTest[,j] <- prediction(test, model, classifier = "rrBLUP");
            r2Classif <- c(r2Classif, rrBLUP=r2(phenoTest, predictionTest[,"rrBLUP"]))},
           {print(paste("unknown classifier, please choose among rf, svm, lasso, rrBLUP"));
             stop()}
      )
  }
  predictionTrain.list <- c(predictionTrain.list, list(predictionTrain))
  predictionTest.list <- c(predictionTest.list, list(predictionTest))
  r2Classif.list <- c(r2Classif.list, list(r2Classif))
}
# aggregate !
switch(method,
       geneticMean={
         aggreg <- aggregateGeneticMean(train=predictionTrain.list, test=predictionTest.list,
                                        target = phenotype, folds=folds)
       },
       dt={
         aggreg <- aggregateDT(train=predictionTrain.list, test=predictionTest.list,
                               target = phenotype, folds=folds)
       },
       lasso={
         aggreg <- aggregateLASSO(train=predictionTrain.list, test=predictionTest.list,
                               target = phenotype, folds=folds)
       },
       rf={
         aggreg <- aggregateRF(train=predictionTrain.list, test=predictionTest.list,
                               target = phenotype, folds=folds)
       },
       # svm, by default
       {aggreg <- aggregateSVM(train=predictionTrain.list, test=predictionTest.list,
                               target = phenotype, folds=folds, kernel=kernel)}
)
# determine best classifier
# first, transform list into a matrix
saveRDS(r2Classif.list, "/Users/nbeaume/Desktop/r2Classif.rds")
r2Classif.list <- t(data.frame(r2Classif.list))
# then, compute the mean r2 for each classifier
meanR2Classif <- apply(r2Classif.list, 2, mean)
# choose the best one
bestClassif <- which.max(meanR2Classif)
# compare aggregation and best classifiers
finalRes <- cbind(bestClassif=r2Classif.list[,bestClassif], aggreg=aggreg,
                  diff=(aggreg-r2Classif.list[,bestClassif]))
print(apply(finalRes, 2, mean))