R/meilleur_lars.R

In CorReg: Linear Regression Based on Linear Structure Between Variables

# 'On estime les coefficients du meilleur mod?le obtenu (meilleur au sens du crit?re choisi)
# '
# '
meilleur_lars <- function(lars = lars, X = X, Y = Y, mode = c("MSE", "BIC"), intercept = TRUE, K = NULL, groupe = NULL, Amax = NULL, OLS = TRUE) {
  mode = match.arg(mode)
  X = as.matrix(X)
  if (is.null(Amax)) {
    Amax = ncol(X) + 1
  }
  if (mode == "BIC") {
    vrais_vect = c()
    coef = as.matrix(coef(lars))
    coefbin = coef
    coefbin[coef != 0] = 1
    coef = as.matrix(coef[rowSums(coefbin) <= Amax, ])
    A = rep(0, times = ncol(X) + intercept)
    if (!is.null(nrow(coef))) {
      for (i in 1:nrow(coef)) {
        qui = which(coef[i, ] != 0)
        vrais_vect[i] = OLS(X = X[, qui], Y = Y, intercept = intercept, Bic = TRUE)$BIC
      }
      qui = which.min(vrais_vect)
      BIC = vrais_vect[qui]
      qui = which(coef[qui, ] != 0)
      A[c(intercept, qui + intercept)] = OLS(X = X[, qui], Y = Y, intercept = intercept)$beta
    }
    
    return(list(A = A, BIC = BIC))
    
  } else { # on regarde les MSE et on prend le meilleur
    K = min(K, nrow(X))
    vrais_vect = c()
    coef = as.matrix(coef(lars))
    coefbin = coef
    coefbin[coef != 0] = 1
    coef = as.matrix(coef[rowSums(coefbin) <= Amax, ])
    A = rep(0, times = ncol(X) + intercept)
    MSE = var(Y)
    if (!is.null(nrow(coef))) {
      for (i in 1:nrow(coef)) {
        qui = which(coef[i, ] != 0)
        vrais_vect[i] = CVMSE(X = X[, qui], intercept = intercept, Y = Y, K = K, groupe = groupe)
      }
      qui = which.min(vrais_vect)
      MSE = vrais_vect[qui]
      qui = which(coef[qui, ] != 0)
      if (length(qui) > 0) {
        A[c(intercept, qui + intercept)] = OLS(X = as.matrix(X[, qui]), Y = Y, intercept = intercept)$beta
      } # sinon ne garde rien donc on laisse A sur sa valeur initiale nulle
    }
    return(list(A = A, CVMSE = MSE))
  }
}