R/LearnerClustKKMeans.R

In mlr3cluster: Cluster Extension for 'mlr3'

#' @title Kernel K-Means Clustering Learner
#'
#' @name mlr_learners_clust.kkmeans
#' @include LearnerClust.R
#' @include aaa.R
#'
#' @description
#' A [LearnerClust] for kernel k-means clustering implemented in [kernlab::kkmeans()].
#' [kernlab::kkmeans()] doesn't have a default value for the number of clusters.
#' Therefore, the `centers` parameter here is set to 2 by default.
#' Kernel parameters have to be passed directly and not by using the `kpar` list in `kkmeans`.
#' The predict method finds the nearest center in kernel distance to
#' assign clusters for new data points.
#'
#' @templateVar id clust.kkmeans
#' @template learner
#' @template example
#'
#' @export
LearnerClustKKMeans = R6Class("LearnerClustKKMeans",
  inherit = LearnerClust,
  public = list(
    #' @description
    #' Creates a new instance of this [R6][R6::R6Class] class.
    initialize = function() {
      ps = ps(
        centers = p_uty(tags = c("required", "train"), default = 2L,
          custom_check = function(x) {
            if (test_data_frame(x)) {
              return(TRUE)
            } else if (test_int(x)) {
              assert_true(x >= 1L)
            } else {
              return("`centers` must be integer or data.frame with initial cluster centers")
            }
          }
        ),
        kernel = p_fct(default = "rbfdot",
          levels = c( "vanilladot", "polydot", "rbfdot", "tanhdot", "laplacedot", "besseldot", "anovadot", "splinedot"),
          tags = "train"),
        sigma = p_dbl(lower = 0, tags = "train"),
        degree = p_int(default = 3L, lower = 1L, tags = "train"),
        scale = p_dbl(default = 1, lower = 0, tags = "train"),
        offset = p_dbl(default = 1, tags = "train"),
        order = p_int(default = 1L, tags = "train"),
        alg = p_fct(levels = c("kkmeans", "kerninghan"), default = "kkmeans", tags = "train"),
        p = p_dbl(default = 1, tags = "train")
      )
      ps$values = list(centers = 2L)

      # add deps
      ps$add_dep(
        "sigma", "kernel",
        CondAnyOf$new(c("rbfdot", "anovadot", "besseldot", "laplacedot")))
      ps$add_dep("degree", "kernel", CondAnyOf$new(c("polydot", "anovadot", "besseldot")))
      ps$add_dep("scale", "kernel", CondAnyOf$new(c("polydot", "tanhdot")))
      ps$add_dep("offset", "kernel", CondAnyOf$new(c("polydot", "tanhdot")))
      ps$add_dep("order", "kernel", CondEqual$new("besseldot"))

      super$initialize(
        id = "clust.kkmeans",
        feature_types = c("logical", "integer", "numeric"),
        predict_types = "partition",
        param_set = ps,
        properties = c("partitional", "exclusive", "complete"),
        packages = "kernlab",
        man = "mlr3cluster::mlr_learners_clust.kkmeans",
        label = "Kernel K-Means"
      )
    }
  ),
  private = list(
    .train = function(task) {
      check_centers_param(self$param_set$values$centers, task, test_data_frame, "centers")

      pv = self$param_set$get_values(tags = "train")
      m = invoke(kernlab::kkmeans, x = as.matrix(task$data()), .args = pv)
      if (self$save_assignments) {
        self$assignments = m[seq_along(m)]
      }
      return(m)
    },
    .predict = function(task) {
      # all of predict is taken from mlr2

      c = kernlab::centers(self$model)
      K = kernlab::kernelf(self$model)

      # kernel product between each new datapoint and the centers
      d_xc = matrix(kernlab::kernelMatrix(K, as.matrix(task$data()), c), ncol = nrow(c))
      # kernel product between each new datapoint and itself: rows are identical
      d_xx = matrix(rep(diag(kernlab::kernelMatrix(K, as.matrix(task$data()))),
                        each = ncol(d_xc)), ncol = ncol(d_xc), byrow = TRUE)
      # kernel product between each center and itself: columns are identical
      d_cc = matrix(rep(diag(kernlab::kernelMatrix(K, as.matrix(c))),
                        each = nrow(d_xc)), nrow = nrow(d_xc))
      # this is the squared kernel distance to the centers
      d2 = d_xx + d_cc - 2 * d_xc
      # the nearest center determines cluster assignment
      partition = apply(d2, 1, function(x) which.min(x))

      PredictionClust$new(task = task, partition = partition)
    }
  )
)

learners[["clust.kkmeans"]] = LearnerClustKKMeans