R/agghoo.R

   1 #' agghoo
   2 #'
   3 #' Run the agghoo procedure. (...)
   4 #'
   5 #' @param data Data frame or matrix containing the data in lines.
   6 #' @param target The target values to predict. Generally a vector.
   7 #' @param task "classification" or "regression". Default:
   8 #'        regression if target is numerical, classification otherwise.
   9 #' @param gmodel A "generic model", which is a function returning a predict
  10 #'        function (taking X as only argument) from the tuple
  11 #'        (dataHO, targetHO, param), where 'HO' stands for 'Hold-Out',
  12 #'        referring to cross-validation. Cross-validation is run on an array
  13 #'        of 'param's. See params argument. Default: see R6::Model.
  14 #' @param params A list of parameters. Often, one list cell is just a
  15 #'        numerical value, but in general it could be of any type.
  16 #'        Default: see R6::Model.
  17 #' @param quality A function assessing the quality of a prediction.
  18 #'        Arguments are y1 and y2 (comparing a prediction to known values).
  19 #'        Default: see R6::AgghooCV.
  20 #'
  21 #' @return An R6::AgghooCV object.
  22 #'
  23 #' @examples
  24 #' # Regression:
  25 #' a_reg <- agghoo(iris[,-c(2,5)], iris[,2])
  26 #' a_reg$fit()
  27 #' pr <- a_reg$predict(iris[,-c(2,5)] + rnorm(450, sd=0.1))
  28 #' # Classification
  29 #' a_cla <- agghoo(iris[,-5], iris[,5])
  30 #' a_cla$fit(mode="standard")
  31 #' pc <- a_cla$predict(iris[,-5] + rnorm(600, sd=0.1))
  32 #'
  33 #' @export
  34 agghoo <- function(data, target, task = NULL, gmodel = NULL, params = NULL, quality = NULL) {
  35     # Args check:
  36   if (!is.data.frame(data) && !is.matrix(data))
  37     stop("data: data.frame or matrix")
  38   if (!is.numeric(target) && !is.factor(target) && !is.character(target))
  39     stop("target: numeric, factor or character vector")
  40   if (!is.null(task))
  41     task = match.arg(task, c("classification", "regression"))
  42   if (is.character(gmodel))
  43     gmodel <- match.arg(gmodel, c("knn", "ppr", "rf", "tree"))
  44   else if (!is.null(gmodel) && !is.function(gmodel))
  45     # No further checks here: fingers crossed :)
  46     stop("gmodel: function(dataHO, targetHO, param) --> function(X) --> y")
  47   if (is.numeric(params) || is.character(params))
  48     params <- as.list(params)
  49   if (!is.list(params) && !is.null(params))
  50     stop("params: numerical, character, or list (passed to model)")
  51   if (is.function(gmodel) && !is.list(params))
  52     stop("params must be provided when using a custom model")
  53   if (is.list(params) && is.null(gmodel))
  54     stop("model (or family) must be provided when using custom params")
  55   if (!is.null(quality) && !is.function(quality))
  56     # No more checks here as well... TODO:?
  57     stop("quality: function(y1, y2) --> Real")
  58
  59   if (is.null(task)) {
  60     if (is.numeric(target))
  61       task = "regression"
  62     else
  63       task = "classification"
  64   }
  65   # Build Model object (= list of parameterized models)
  66   model <- Model$new(data, target, task, gmodel, params)
  67   # Return AgghooCV object, to run and predict
  68   AgghooCV$new(data, target, task, model, quality)
  69 }
  70
  71 #' compareToStandard
  72 #'
  73 #' Temporary function to compare agghoo to CV
  74 #' (TODO: extended, in another file, more tests - when faster code).
  75 #'
  76 #' @export
  77 compareToStandard <- function(df, t_idx, task = NULL, rseed = -1) {
  78   if (rseed >= 0)
  79     set.seed(rseed)
  80   if (is.null(task))
  81     task <- ifelse(is.numeric(df[,t_idx]), "regression", "classification")
  82   n <- nrow(df)
  83   test_indices <- sample( n, round(n / ifelse(n >= 500, 10, 5)) )
  84   a <- agghoo(df[-test_indices,-t_idx], df[-test_indices,t_idx], task)
  85   a$fit(mode="agghoo") #default mode
  86   pa <- a$predict(df[test_indices,-t_idx])
  87   print(paste("error agghoo",
  88               ifelse(task == "classification",
  89                      mean(p != df[test_indices,t_idx]),
  90                      mean(abs(pa - df[test_indices,t_idx])))))
  91   # Compare with standard cross-validation:
  92   a$fit(mode="standard")
  93   ps <- a$predict(df[test_indices,-t_idx])
  94   print(paste("error CV",
  95               ifelse(task == "classification",
  96                      mean(ps != df[test_indices,t_idx]),
  97                      mean(abs(ps - df[test_indices,t_idx])))))
  98 }